Strategy Ranking · BTC Trading AI

🟢 En producción (resaltadas)

no shorts (V66 canonical)70,576.0% compound · α +239.2

canonical V66 (full notional)70,576.0% compound · α +239.2

V66 canonical (no pyramid) — bugfix engine42,636.5% compound · α +192.8

V66 alone (bugfix engine)42,636.5% compound · α +192.8

V66 canonical— compound · α —

V66 alone bugfix— compound · α —

Por qué los números altos arriba ≠ candidatos reales: rondas pre-R173 (R130-R172) corren bajo el engine viejo con intra-bar lookahead + apalancamiento sin coste. R154 muestra +250K compound pero es la MISMA estrategia R151-A bajo bug-fixed engine que da +113K (R174). El número alto es artefacto. Por eso los CANDIDATOS reales (azul) son R174+. Filtra "ocultar buggy-engine" si quieres ver solo resultados honestos.

Round Variant Compound % Min α Beats BH Trades WR OOT α

R154 skip shorts if 21d return > +15%BUGGY-ENGINE 250,538% +245.1 4/4 88 sh — —

⚠️ BUGGY-ENGINE — Esta ronda es pre-R173 (engine con bug intra-bar lookahead). El compound está inflado vs realidad de producción. Para R151-A, la versión bug-fixed honesta es R174 (+113,981%). Por eso los buggy aparecen arriba pero los candidatos reales son los R173+.

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	1,055.4%	+951.4%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=88 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.15

shorts_skipped

243

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Same as R151 + regime filter: skip shorts if recent N-day return > threshold (variants: 7d>+5/+10, 14d>+10/+15, 21d>+15/+20).

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r154_regime_filtered_shorts.py

Experimento: R154: regime-filtered shorts (R151-A base)
Origen: outputs/round_154_regime_shorts/metadata.json · → ver detalle de la ronda

R154 skip shorts if 7d return > +10%BUGGY-ENGINE 234,511% +245.1 4/4 90 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	981.5%	+877.5%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=90 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.1

shorts_skipped

160

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Same as R151 + regime filter: skip shorts if recent N-day return > threshold (variants: 7d>+5/+10, 14d>+10/+15, 21d>+15/+20).

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r154_regime_filtered_shorts.py

Experimento: R154: regime-filtered shorts (R151-A base)
Origen: outputs/round_154_regime_shorts/metadata.json · → ver detalle de la ronda

R154 skip shorts if 21d return > +20%BUGGY-ENGINE 213,591% +245.1 4/4 89 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	885.1%	+781.1%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=89 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.2

shorts_skipped

219

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Same as R151 + regime filter: skip shorts if recent N-day return > threshold (variants: 7d>+5/+10, 14d>+10/+15, 21d>+15/+20).

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r154_regime_filtered_shorts.py

Experimento: R154: regime-filtered shorts (R151-A base)
Origen: outputs/round_154_regime_shorts/metadata.json · → ver detalle de la ronda

R151 short P(SL)>0.65, exit P(TP)>0.35, sl=8%, max=40hBUGGY-ENGINE 212,591% +245.1 4/4 93 sh 56.6% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	880.5%	+776.5%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=56.550925925925924%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 longs (canonical) + SHORT module: cuando V66 está en cash, si P(SL)>0.65 (3/5 seeds), abre short. Cierra en P(TP)>0.35, stop 8%, o 40h timeout.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r151_shorts_in_gaps.py

Experimento: R151: V66 long + SHORT during cash gaps
Origen: outputs/round_151_shorts_gaps/metadata.json · → ver detalle de la ronda

R154 R151-A (no regime filter)BUGGY-ENGINE 212,591% +245.1 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	880.5%	+776.5%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

shorts_skipped

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Same as R151 + regime filter: skip shorts if recent N-day return > threshold (variants: 7d>+5/+10, 14d>+10/+15, 21d>+15/+20).

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r154_regime_filtered_shorts.py

Experimento: R154: regime-filtered shorts (R151-A base)
Origen: outputs/round_154_regime_shorts/metadata.json · → ver detalle de la ronda

R160 R151-A no friction (baseline)BUGGY-ENGINE 212,591% +245.1 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	880.5%	+776.5%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

funding

0.0

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 cascade: entry on regime-specific thresholds (bull/lateral/bear via 20d slope) → exit on (init_sl 10% OR trail ratchet 22→18→5% by unrealized OR DRSI>78 OR 3/5 GRUs danger OR peak_drop of avg_tp_bar) → cooldown 4/48 post-any-exit/post-loss

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r160_funding_stress.py

Experimento: R160: friction stress on R151-A
Origen: outputs/round_160_friction_stress/metadata.json · → ver detalle de la ronda

R160 + funding 0.01%/8hBUGGY-ENGINE 191,908% +234.1 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	834.7%	+730.7%	104%	—	—
F1 Bull (2019-21)	845.4%	-316.6%	1162%	—	—
F2 Recovery (2022-24)	468.5%	+234.5%	234%	—	—
F3 Lateral (2024-25)	282.2%	+282.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

funding

0.0001

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r160_funding_stress.py

Experimento: R160: friction stress on R151-A
Origen: outputs/round_160_friction_stress/metadata.json · → ver detalle de la ronda

R154 skip shorts if 7d return > +5%BUGGY-ENGINE 181,144% +245.1 4/4 88 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	735.5%	+631.5%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=88 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.05

shorts_skipped

254

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Same as R151 + regime filter: skip shorts if recent N-day return > threshold (variants: 7d>+5/+10, 14d>+10/+15, 21d>+15/+20).

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r154_regime_filtered_shorts.py

Experimento: R154: regime-filtered shorts (R151-A base)
Origen: outputs/round_154_regime_shorts/metadata.json · → ver detalle de la ronda

R137 R117 MACD_div local reproductionBUGGY-ENGINE 179,453% -42.0 3/4 6,544 — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	+480.5%	104%	—	—
F1 Bull (2019-21)	—	+2,405.6%	1162%	—	—
F2 Recovery (2022-24)	—	-42.0%	234%	—	—
F3 Lateral (2024-25)	—	+270.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=6544 · WR=—

Métricas y config crudas (metadata.json)

folds_trades

[1658, 1624, 1639, 1623]

folds_wr

[44.27020506634499, 45.19704433497537, 44.05125076266016,...

folds_mdd

[23.591372375173027, 18.580826219207594, 19.8152419508006...

bh_per_fold

[59.18068329501152, 99.98283255066114, 228.4183620424153,...

reproduces_r117

True

compound_diff_pct

1.591644913315314e-05

r117_expected

{"compound": 179453.2, "min_alpha": -42.0, "beats_bh": 3,...

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r137_macd_repro.py

Experimento: R117 MACD_div local reproduction
Origen: outputs/round_137/metadata.json · → ver detalle de la ronda

R137 R117 MACD_div local reproductionBUGGY-ENGINE 179,453% -42.0 3/4 6,544 — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	+480.5%	104%	—	—
F1 Bull (2019-21)	—	+2,405.6%	1162%	—	—
F2 Recovery (2022-24)	—	-42.0%	234%	—	—
F3 Lateral (2024-25)	—	+270.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=6544 · WR=—

Métricas y config crudas (metadata.json)

folds_trades

[1658, 1624, 1639, 1623]

folds_wr

[44.27020506634499, 45.19704433497537, 44.05125076266016,...

folds_mdd

[23.591372375173027, 18.580826219207594, 19.8152419508006...

bh_per_fold

[59.18068329501152, 99.98283255066114, 228.4183620424153,...

reproduces_r117

True

compound_diff_pct

1.591644913315314e-05

r117_expected

{"compound": 179453.2, "min_alpha": -42.0, "beats_bh": 3,...

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r137_macd_repro.py

Experimento: R117 MACD_div local reproduction
Origen: outputs/round_137_macd_repro/metadata.json · → ver detalle de la ronda

R154 skip shorts if 14d return > +10%BUGGY-ENGINE 172,844% +245.1 4/4 91 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	697.3%	+593.3%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=91 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.1

shorts_skipped

153

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Same as R151 + regime filter: skip shorts if recent N-day return > threshold (variants: 7d>+5/+10, 14d>+10/+15, 21d>+15/+20).

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r154_regime_filtered_shorts.py

Experimento: R154: regime-filtered shorts (R151-A base)
Origen: outputs/round_154_regime_shorts/metadata.json · → ver detalle de la ronda

R154 skip shorts if 14d return > +15%BUGGY-ENGINE 170,820% +245.1 4/4 91 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	687.9%	+583.9%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=91 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.15

shorts_skipped

109

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Same as R151 + regime filter: skip shorts if recent N-day return > threshold (variants: 7d>+5/+10, 14d>+10/+15, 21d>+15/+20).

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r154_regime_filtered_shorts.py

Experimento: R154: regime-filtered shorts (R151-A base)
Origen: outputs/round_154_regime_shorts/metadata.json · → ver detalle de la ronda

R160 + funding 0.03%/8h (worst case)BUGGY-ENGINE 156,375% +212.8 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	749.5%	+645.5%	104%	—	—
F1 Bull (2019-21)	831.3%	-330.7%	1162%	—	—
F2 Recovery (2022-24)	447.2%	+213.2%	234%	—	—
F3 Lateral (2024-25)	261.5%	+261.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

funding

0.0003

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r160_funding_stress.py

Experimento: R160: friction stress on R151-A
Origen: outputs/round_160_friction_stress/metadata.json · → ver detalle de la ronda

R160 + regime slippageBUGGY-ENGINE 143,856% +198.3 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	812.6%	+708.6%	104%	—	—
F1 Bull (2019-21)	738.9%	-423.1%	1162%	—	—
F2 Recovery (2022-24)	432.7%	+198.7%	234%	—	—
F3 Lateral (2024-25)	253.0%	+253.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

funding

0.0

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r160_funding_stress.py

Experimento: R160: friction stress on R151-A
Origen: outputs/round_160_friction_stress/metadata.json · → ver detalle de la ronda

R181 skip shorts if 21d return > +15% 136,242% +193.8 4/4 89 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	698.2%	+641.5%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=89 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.15

shorts_skipped

243

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + regime filter (skip shorts when N-day return > threshold). 7 variants tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r181_regime_filtered_honest.py

Experimento: R181: regime-filtered R151-A under bug-fixed engine
Origen: outputs/round_181_regime_filtered_honest/metadata.json · → ver detalle de la ronda

R160 + funding 0.01% + regime slipBUGGY-ENGINE 129,857% +188.2 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	770.0%	+666.0%	104%	—	—
F1 Bull (2019-21)	732.6%	-429.4%	1162%	—	—
F2 Recovery (2022-24)	422.6%	+188.6%	234%	—	—
F3 Lateral (2024-25)	243.3%	+243.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

funding

0.0001

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r160_funding_stress.py

Experimento: R160: friction stress on R151-A
Origen: outputs/round_160_friction_stress/metadata.json · → ver detalle de la ronda

R181 skip shorts if 7d return > +10% 124,625% +193.8 4/4 91 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	630.2%	+573.5%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=91 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.1

shorts_skipped

160

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + regime filter (skip shorts when N-day return > threshold). 7 variants tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r181_regime_filtered_honest.py

Experimento: R181: regime-filtered R151-A under bug-fixed engine
Origen: outputs/round_181_regime_filtered_honest/metadata.json · → ver detalle de la ronda

R181 skip shorts if 21d return > +20% 115,235% +193.8 4/4 90 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	575.3%	+518.6%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=90 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.2

shorts_skipped

219

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + regime filter (skip shorts when N-day return > threshold). 7 variants tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r181_regime_filtered_honest.py

Experimento: R181: regime-filtered R151-A under bug-fixed engine
Origen: outputs/round_181_regime_filtered_honest/metadata.json · → ver detalle de la ronda

R151 short P(SL)>0.70, exit P(TP)>0.30, sl=10%, max=40hBUGGY-ENGINE 114,349% +321.6 4/4 45 sh 73.6% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	384.7%	+280.7%	104%	—	—
F1 Bull (2019-21)	629.5%	-532.5%	1162%	—	—
F2 Recovery (2022-24)	657.3%	+423.3%	234%	—	—
F3 Lateral (2024-25)	327.4%	+327.4%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=45 · WR=73.61111111111111%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 longs (canonical) + SHORT module: cuando V66 está en cash, si P(SL)>0.65 (3/5 seeds), abre short. Cierra en P(TP)>0.35, stop 8%, o 40h timeout.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r151_shorts_in_gaps.py

Experimento: R151: V66 long + SHORT during cash gaps
Origen: outputs/round_151_shorts_gaps/metadata.json · → ver detalle de la ronda

R174 R151-A shorts (bugfix engine)CANDIDATE 113,981% +193.8 4/4 94 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	567.9%	+463.9%	104%	—	—
F1 Bull (2019-21)	853.9%	-308.1%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+194.2%	234%	—	—
F3 Lateral (2024-25)	239.0%	+239.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r174_r151_bugfix.py

Experimento: R174: R151-A re-validated under bug-fixed engine
Origen: outputs/round_174_r151_bugfix/metadata.json · → ver detalle de la ronda

R181 R151-A base (no filter, honest engine)CANDIDATE 113,981% +193.8 4/4 94 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	567.9%	+511.2%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=—

Métricas y config crudas (metadata.json)

shorts_skipped

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + regime filter (skip shorts when N-day return > threshold). 7 variants tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r181_regime_filtered_honest.py

Experimento: R181: regime-filtered R151-A under bug-fixed engine
Origen: outputs/round_181_regime_filtered_honest/metadata.json · → ver detalle de la ronda

R182 R151-A base (full notional)CANDIDATE 113,981% +193.8 4/4 94 sh 54.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	567.9%	+511.2%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=54.02777777777777%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + ADAPTIVE Kelly sizing based on rolling short WR (last N shorts). 4 schedules + half-Kelly fixed tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r182_adaptive_kelly.py

Experimento: R182: R151-A with adaptive Kelly sizing based on rolling short WR
Origen: outputs/round_182_adaptive_kelly/metadata.json · → ver detalle de la ronda

R181 skip shorts if 7d return > +5% 110,320% +193.8 4/4 89 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	546.5%	+489.8%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=89 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.05

shorts_skipped

254

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + regime filter (skip shorts when N-day return > threshold). 7 variants tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r181_regime_filtered_honest.py

Experimento: R181: regime-filtered R151-A under bug-fixed engine
Origen: outputs/round_181_regime_filtered_honest/metadata.json · → ver detalle de la ronda

R181 skip shorts if 14d return > +10% 97,054% +193.8 4/4 91 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	468.8%	+412.1%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=91 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.1

shorts_skipped

153

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + regime filter (skip shorts when N-day return > threshold). 7 variants tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r181_regime_filtered_honest.py

Experimento: R181: regime-filtered R151-A under bug-fixed engine
Origen: outputs/round_181_regime_filtered_honest/metadata.json · → ver detalle de la ronda

R160 + all + half-Kelly shorts (50%)BUGGY-ENGINE 89,759% +182.9 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	654.0%	+550.0%	104%	—	—
F1 Bull (2019-21)	631.2%	-530.8%	1162%	—	—
F2 Recovery (2022-24)	417.3%	+183.3%	234%	—	—
F3 Lateral (2024-25)	215.1%	+215.1%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

funding

0.0001

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r160_funding_stress.py

Experimento: R160: friction stress on R151-A
Origen: outputs/round_160_friction_stress/metadata.json · → ver detalle de la ronda

R181 skip shorts if 14d return > +15% 88,844% +193.8 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	420.7%	+364.0%	104%	—	—
F1 Bull (2019-21)	853.9%	+755.9%	1162%	—	—
F2 Recovery (2022-24)	428.2%	+193.8%	234%	—	—
F3 Lateral (2024-25)	239.0%	+233.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

lookback

return_th

0.15

shorts_skipped

109

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + regime filter (skip shorts when N-day return > threshold). 7 variants tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r181_regime_filtered_honest.py

Experimento: R181: regime-filtered R151-A under bug-fixed engine
Origen: outputs/round_181_regime_filtered_honest/metadata.json · → ver detalle de la ronda

R182 Adaptive binary (WR last 10, full/30%) 77,859% +157.7 4/4 94 sh 54.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	521.2%	+464.5%	104%	—	—
F1 Bull (2019-21)	713.7%	+615.7%	1162%	—	—
F2 Recovery (2022-24)	392.1%	+157.7%	234%	—	—
F3 Lateral (2024-25)	213.4%	+207.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=54.02777777777777%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + ADAPTIVE Kelly sizing based on rolling short WR (last N shorts). 4 schedules + half-Kelly fixed tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r182_adaptive_kelly.py

Experimento: R182: R151-A with adaptive Kelly sizing based on rolling short WR
Origen: outputs/round_182_adaptive_kelly/metadata.json · → ver detalle de la ronda

R174 R151-A + half-Kelly (bugfix)CANDIDATE 76,873% +178.4 4/4 94 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	482.3%	+378.3%	104%	—	—
F1 Bull (2019-21)	732.8%	-429.2%	1162%	—	—
F2 Recovery (2022-24)	412.8%	+178.8%	234%	—	—
F3 Lateral (2024-25)	209.5%	+209.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r174_r151_bugfix.py

Experimento: R174: R151-A re-validated under bug-fixed engine
Origen: outputs/round_174_r151_bugfix/metadata.json · → ver detalle de la ronda

R182 Half-Kelly fixed (size=0.5 always) 76,873% +178.4 4/4 94 sh 54.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	482.3%	+425.6%	104%	—	—
F1 Bull (2019-21)	732.8%	+634.8%	1162%	—	—
F2 Recovery (2022-24)	412.8%	+178.4%	234%	—	—
F3 Lateral (2024-25)	209.5%	+203.7%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=54.02777777777777%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + ADAPTIVE Kelly sizing based on rolling short WR (last N shorts). 4 schedules + half-Kelly fixed tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r182_adaptive_kelly.py

Experimento: R182: R151-A with adaptive Kelly sizing based on rolling short WR
Origen: outputs/round_182_adaptive_kelly/metadata.json · → ver detalle de la ronda

R182 Adaptive Kelly (WR last 10, kelly schedule) 72,736% +154.6 4/4 94 sh 54.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	498.7%	+442.0%	104%	—	—
F1 Bull (2019-21)	695.8%	+597.8%	1162%	—	—
F2 Recovery (2022-24)	389.0%	+154.6%	234%	—	—
F3 Lateral (2024-25)	212.6%	+206.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=54.02777777777777%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + ADAPTIVE Kelly sizing based on rolling short WR (last N shorts). 4 schedules + half-Kelly fixed tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r182_adaptive_kelly.py

Experimento: R182: R151-A with adaptive Kelly sizing based on rolling short WR
Origen: outputs/round_182_adaptive_kelly/metadata.json · → ver detalle de la ronda

R182 Adaptive Kelly (WR last 20, kelly schedule) 72,044% +159.7 4/4 94 sh 54.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	503.3%	+446.6%	104%	—	—
F1 Bull (2019-21)	695.8%	+597.8%	1162%	—	—
F2 Recovery (2022-24)	394.1%	+159.7%	234%	—	—
F3 Lateral (2024-25)	204.1%	+198.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=54.02777777777777%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + ADAPTIVE Kelly sizing based on rolling short WR (last N shorts). 4 schedules + half-Kelly fixed tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r182_adaptive_kelly.py

Experimento: R182: R151-A with adaptive Kelly sizing based on rolling short WR
Origen: outputs/round_182_adaptive_kelly/metadata.json · → ver detalle de la ronda

R110 R110 local reproduction 70,576% +239.2 —/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+378.4%	104%	—	—
F1 Bull (2019-21)	537.3%	+439.3%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.2%	234%	—	—
F3 Lateral (2024-25)	261.3%	+255.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

expected_compound

70576

expected_min_alpha

245

reproduces

True

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r110_local.py

Experimento: R110 local reproduction
Origen: outputs/round_110_repro/metadata.json · → ver detalle de la ronda

R136 R110 local reproductionBUGGY-ENGINE 70,576% +239.2 —/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+378.4%	104%	—	—
F1 Bull (2019-21)	537.3%	+439.3%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.2%	234%	—	—
F3 Lateral (2024-25)	261.3%	+255.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

expected_compound

70576

expected_min_alpha

245

reproduces

True

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Experimento: R110 local reproduction
Origen: outputs/round_136/metadata.json · → ver detalle de la ronda

R146 canonical (10% fixed)BUGGY-ENGINE 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

27.672150676086755

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r146_atr_stops.py

Experimento: R146: ATR-adaptive initial SL on canonical V66
Origen: outputs/round_146_atr_stops/metadata.json · → ver detalle de la ronda

R148 (unnamed)BUGGY-ENGINE 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

tau

0.0

trades_total

300

skip_pct

0.0

avg_mdd

27.672150676086755

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r148_meta_label.py

Experimento: R148: meta-labeling V66 + R134 P(v66_zone) filter
Origen: outputs/round_148_metalabel/metadata.json · → ver detalle de la ronda

R150 canonical (no breaker)BUGGY-ENGINE 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

27.672150676086755

circuit_triggers

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r150_maxdd_circuit.py

Experimento: R150: Max-DD circuit breaker on V66
Origen: outputs/round_150_maxdd/metadata.json · → ver detalle de la ronda

R151 no shorts (V66 canonical)PRODBUGGY-ENGINE 70,576% +239.2 4/4 — 0.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 longs (canonical) + SHORT module: cuando V66 está en cash, si P(SL)>0.65 (3/5 seeds), abre short. Cierra en P(TP)>0.35, stop 8%, o 40h timeout.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r151_shorts_in_gaps.py

Experimento: R151: V66 long + SHORT during cash gaps
Origen: outputs/round_151_shorts_gaps/metadata.json · → ver detalle de la ronda

R151 short P(SL)>0.80, exit P(TP)>0.20, sl=15%, max=40hBUGGY-ENGINE 70,576% +239.2 4/4 — 0.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 longs (canonical) + SHORT module: cuando V66 está en cash, si P(SL)>0.65 (3/5 seeds), abre short. Cierra en P(TP)>0.35, stop 8%, o 40h timeout.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r151_shorts_in_gaps.py

Experimento: R151: V66 long + SHORT during cash gaps
Origen: outputs/round_151_shorts_gaps/metadata.json · → ver detalle de la ronda

R156 R151-A (canonical R069 for shorts) [baseline]BUGGY-ENGINE 70,576% +239.2 4/4 — 0.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier asymmetric: k_up=7, k_dn=10 → favors P(SL) responsiveness

Entrada / salida

Dual-model: V66 longs use canonical GRU; shorts use R155 short-specific model

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r156_dual_model.py

Experimento: R156: V66 canonical longs + R155 short-trained shorts
Origen: outputs/round_156_dual_model/metadata.json · → ver detalle de la ronda

R156 R155 P(SL)>0.65, exit P(TP)>0.35, sl=8%BUGGY-ENGINE 70,576% +239.2 4/4 — 0.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier asymmetric: k_up=7, k_dn=10 → favors P(SL) responsiveness

Entrada / salida

Dual-model: V66 longs use canonical GRU; shorts use R155 short-specific model

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r156_dual_model.py

Experimento: R156: V66 canonical longs + R155 short-trained shorts
Origen: outputs/round_156_dual_model/metadata.json · → ver detalle de la ronda

R162 canonical V66 (full notional)PRODBUGGY-ENGINE 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r162_r134_size_multiplier.py

Experimento: R162: R134 P(v66_zone) as position size multiplier
Origen: outputs/round_162_r134_size_mult/metadata.json · → ver detalle de la ronda

R156 R155 P(SL)>0.60, exit P(TP)>0.30, sl=8%BUGGY-ENGINE 70,209% +239.2 4/4 11 sh 54.5% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	432.3%	+328.3%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=11 · WR=54.54545454545454%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier asymmetric: k_up=7, k_dn=10 → favors P(SL) responsiveness

Entrada / salida

Dual-model: V66 longs use canonical GRU; shorts use R155 short-specific model

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r156_dual_model.py

Experimento: R156: V66 canonical longs + R155 short-trained shorts
Origen: outputs/round_156_dual_model/metadata.json · → ver detalle de la ronda

R182 Adaptive Kelly (WR last 5, kelly schedule) 62,842% +157.9 4/4 94 sh 54.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	407.6%	+350.9%	104%	—	—
F1 Bull (2019-21)	734.2%	+636.2%	1162%	—	—
F2 Recovery (2022-24)	392.3%	+157.9%	234%	—	—
F3 Lateral (2024-25)	201.9%	+196.1%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=94 · WR=54.02777777777777%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A under bug-fixed engine + ADAPTIVE Kelly sizing based on rolling short WR (last N shorts). 4 schedules + half-Kelly fixed tested.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r182_adaptive_kelly.py

Experimento: R182: R151-A with adaptive Kelly sizing based on rolling short WR
Origen: outputs/round_182_adaptive_kelly/metadata.json · → ver detalle de la ronda

R174 R151-B more conservative (bugfix)CANDIDATE 54,068% +125.3 4/4 46 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	182.0%	+78.0%	104%	—	—
F1 Bull (2019-21)	616.7%	-545.3%	1162%	—	—
F2 Recovery (2022-24)	568.3%	+334.3%	234%	—	—
F3 Lateral (2024-25)	301.1%	+301.1%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=46 · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r174_r151_bugfix.py

Experimento: R174: R151-A re-validated under bug-fixed engine
Origen: outputs/round_174_r151_bugfix/metadata.json · → ver detalle de la ronda

R150 DD 30% / pause 30dBUGGY-ENGINE 53,423% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	382.6%	-779.4%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

25.739174428102647

circuit_triggers

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r150_maxdd_circuit.py

Experimento: R150: Max-DD circuit breaker on V66
Origen: outputs/round_150_maxdd/metadata.json · → ver detalle de la ronda

R156 R155 P(SL)>0.55, exit P(TP)>0.35, sl=8%BUGGY-ENGINE 52,673% +239.2 4/4 185 sh 48.7% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	301.7%	+197.7%	104%	—	—
F1 Bull (2019-21)	534.0%	-628.0%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=185 · WR=48.728813559322035%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier asymmetric: k_up=7, k_dn=10 → favors P(SL) responsiveness

Entrada / salida

Dual-model: V66 longs use canonical GRU; shorts use R155 short-specific model

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r156_dual_model.py

Experimento: R156: V66 canonical longs + R155 short-trained shorts
Origen: outputs/round_156_dual_model/metadata.json · → ver detalle de la ronda

R151 short P(SL)>0.75, exit P(TP)>0.25, sl=12%, max=40hBUGGY-ENGINE 52,487% +199.3 4/4 9 sh 68.8% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	256.0%	+152.0%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	541.5%	+307.5%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=9 · WR=68.75%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 longs (canonical) + SHORT module: cuando V66 está en cash, si P(SL)>0.65 (3/5 seeds), abre short. Cierra en P(TP)>0.35, stop 8%, o 40h timeout.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r151_shorts_in_gaps.py

Experimento: R151: V66 long + SHORT during cash gaps
Origen: outputs/round_151_shorts_gaps/metadata.json · → ver detalle de la ronda

R146 ATR 2.5x floor 7%BUGGY-ENGINE 48,939% +230.4 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	325.5%	+221.5%	104%	—	—
F1 Bull (2019-21)	505.8%	-656.2%	1162%	—	—
F2 Recovery (2022-24)	464.8%	+230.8%	234%	—	—
F3 Lateral (2024-25)	236.8%	+236.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

32.082204389008545

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r146_atr_stops.py

Experimento: R146: ATR-adaptive initial SL on canonical V66
Origen: outputs/round_146_atr_stops/metadata.json · → ver detalle de la ronda

R156 R155 P(SL)>0.50, exit P(TP)>0.40, sl=8%BUGGY-ENGINE 43,030% +168.3 4/4 487 sh 36.6% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	225.0%	+121.0%	104%	—	—
F1 Bull (2019-21)	545.0%	-617.0%	1162%	—	—
F2 Recovery (2022-24)	469.4%	+235.4%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=487 · WR=36.56055341659624%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier asymmetric: k_up=7, k_dn=10 → favors P(SL) responsiveness

Entrada / salida

Dual-model: V66 longs use canonical GRU; shorts use R155 short-specific model

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r156_dual_model.py

Experimento: R156: V66 canonical longs + R155 short-trained shorts
Origen: outputs/round_156_dual_model/metadata.json · → ver detalle de la ronda

R156 R155 P(SL)>0.50, exit P(TP)>0.40, sl=10%BUGGY-ENGINE 43,030% +168.3 4/4 487 sh 36.6% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	225.0%	+121.0%	104%	—	—
F1 Bull (2019-21)	545.0%	-617.0%	1162%	—	—
F2 Recovery (2022-24)	469.4%	+235.4%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=487 · WR=36.56055341659624%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier asymmetric: k_up=7, k_dn=10 → favors P(SL) responsiveness

Entrada / salida

Dual-model: V66 longs use canonical GRU; shorts use R155 short-specific model

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r156_dual_model.py

Experimento: R156: V66 canonical longs + R155 short-trained shorts
Origen: outputs/round_156_dual_model/metadata.json · → ver detalle de la ronda

R173 V66 canonical (no pyramid) — bugfix enginePROD 42,636% +192.8 4/4 302 56.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	275.0%	+171.0%	104%	—	—
F1 Bull (2019-21)	525.2%	-636.8%	1162%	—	—
F2 Recovery (2022-24)	427.2%	+193.2%	234%	—	—
F3 Lateral (2024-25)	245.8%	+245.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=302 · WR=55.960264900662246%

Métricas y config crudas (metadata.json)

pyramid_adds

pyramid_blocked

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 canonical UNDER BUG-FIXED ENGINE: detect@close[i] → execute@open[i+1]. Pyramid disabled. Cap properly tracked. THIS IS THE HONEST V66 BASELINE.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r173_bugfix.py

Experimento: R173: bug-fixed engine for V66 + pyramid
Origen: outputs/round_173_bugfix/metadata.json · → ver detalle de la ronda

R173 R170-B (0.12/0.65/1.75x/24h) — bugfix engine 42,636% +192.8 4/4 302 56.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	275.0%	+171.0%	104%	—	—
F1 Bull (2019-21)	525.2%	-636.8%	1162%	—	—
F2 Recovery (2022-24)	427.2%	+193.2%	234%	—	—
F3 Lateral (2024-25)	245.8%	+245.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=302 · WR=55.960264900662246%

Métricas y config crudas (metadata.json)

pyramid_adds

pyramid_blocked

517

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 canonical UNDER BUG-FIXED ENGINE: detect@close[i] → execute@open[i+1]. Pyramid disabled. Cap properly tracked. THIS IS THE HONEST V66 BASELINE.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r173_bugfix.py

Experimento: R173: bug-fixed engine for V66 + pyramid
Origen: outputs/round_173_bugfix/metadata.json · → ver detalle de la ronda

R173 R163 baseline (0.10/0.50/1.5x/24h) — bugfix engine 42,636% +192.8 4/4 302 56.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	275.0%	+171.0%	104%	—	—
F1 Bull (2019-21)	525.2%	-636.8%	1162%	—	—
F2 Recovery (2022-24)	427.2%	+193.2%	234%	—	—
F3 Lateral (2024-25)	245.8%	+245.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=302 · WR=55.960264900662246%

Métricas y config crudas (metadata.json)

pyramid_adds

pyramid_blocked

583

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 canonical UNDER BUG-FIXED ENGINE: detect@close[i] → execute@open[i+1]. Pyramid disabled. Cap properly tracked. THIS IS THE HONEST V66 BASELINE.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r173_bugfix.py

Experimento: R173: bug-fixed engine for V66 + pyramid
Origen: outputs/round_173_bugfix/metadata.json · → ver detalle de la ronda

R173 More conservative (0.10/0.30/1.3x/24h) — bugfix 42,636% +192.8 4/4 302 56.0% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	275.0%	+171.0%	104%	—	—
F1 Bull (2019-21)	525.2%	-636.8%	1162%	—	—
F2 Recovery (2022-24)	427.2%	+193.2%	234%	—	—
F3 Lateral (2024-25)	245.8%	+245.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=302 · WR=55.960264900662246%

Métricas y config crudas (metadata.json)

pyramid_adds

pyramid_blocked

388

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 canonical UNDER BUG-FIXED ENGINE: detect@close[i] → execute@open[i+1]. Pyramid disabled. Cap properly tracked. THIS IS THE HONEST V66 BASELINE.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r173_bugfix.py

Experimento: R173: bug-fixed engine for V66 + pyramid
Origen: outputs/round_173_bugfix/metadata.json · → ver detalle de la ronda

R174 V66 alone (bugfix engine)PROD 42,636% +192.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	275.0%	+171.0%	104%	—	—
F1 Bull (2019-21)	525.2%	-636.8%	1162%	—	—
F2 Recovery (2022-24)	427.2%	+193.2%	234%	—	—
F3 Lateral (2024-25)	245.8%	+245.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r174_r151_bugfix.py

Experimento: R174: R151-A re-validated under bug-fixed engine
Origen: outputs/round_174_r151_bugfix/metadata.json · → ver detalle de la ronda

R146 ATR 3x floor 5%BUGGY-ENGINE 29,029% +195.1 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	293.8%	+189.8%	104%	—	—
F1 Bull (2019-21)	293.1%	-868.9%	1162%	—	—
F2 Recovery (2022-24)	474.4%	+240.4%	234%	—	—
F3 Lateral (2024-25)	227.7%	+227.7%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

29.14433857637885

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r146_atr_stops.py

Experimento: R146: ATR-adaptive initial SL on canonical V66
Origen: outputs/round_146_atr_stops/metadata.json · → ver detalle de la ronda

R146 ATR 2x floor 5%BUGGY-ENGINE 28,065% +200.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	275.3%	+171.3%	104%	—	—
F1 Bull (2019-21)	298.8%	-863.2%	1162%	—	—
F2 Recovery (2022-24)	474.4%	+240.4%	234%	—	—
F3 Lateral (2024-25)	227.7%	+227.7%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

30.555060552317062

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r146_atr_stops.py

Experimento: R146: ATR-adaptive initial SL on canonical V66
Origen: outputs/round_146_atr_stops/metadata.json · → ver detalle de la ronda

R146 ATR 2x capped 5-15%BUGGY-ENGINE 28,065% +200.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	275.3%	+171.3%	104%	—	—
F1 Bull (2019-21)	298.8%	-863.2%	1162%	—	—
F2 Recovery (2022-24)	474.4%	+240.4%	234%	—	—
F3 Lateral (2024-25)	227.7%	+227.7%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

30.555060552317062

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r146_atr_stops.py

Experimento: R146: ATR-adaptive initial SL on canonical V66
Origen: outputs/round_146_atr_stops/metadata.json · → ver detalle de la ronda

R150 DD 25% / pause 30dBUGGY-ENGINE 25,534% +189.7 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	246.4%	+142.4%	104%	—	—
F1 Bull (2019-21)	312.6%	-849.4%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	212.7%	+212.7%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

24.18901680121595

circuit_triggers

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r150_maxdd_circuit.py

Experimento: R150: Max-DD circuit breaker on V66
Origen: outputs/round_150_maxdd/metadata.json · → ver detalle de la ronda

R150 DD 20% / pause 15dBUGGY-ENGINE 20,674% +73.9 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	295.1%	+191.1%	104%	—	—
F1 Bull (2019-21)	171.9%	-990.1%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	237.0%	+237.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

20.213634826794753

circuit_triggers

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r150_maxdd_circuit.py

Experimento: R150: Max-DD circuit breaker on V66
Origen: outputs/round_150_maxdd/metadata.json · → ver detalle de la ronda

R150 DD 20% / pause 60dBUGGY-ENGINE 15,747% +113.5 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	204.2%	+100.2%	104%	—	—
F1 Bull (2019-21)	211.5%	-950.5%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	191.5%	+191.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

19.974562037134554

circuit_triggers

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r150_maxdd_circuit.py

Experimento: R150: Max-DD circuit breaker on V66
Origen: outputs/round_150_maxdd/metadata.json · → ver detalle de la ronda

R146 ATR 1.5x capped 4-12%BUGGY-ENGINE 14,737% +69.0 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	237.2%	+133.2%	104%	—	—
F1 Bull (2019-21)	167.0%	-995.0%	1162%	—	—
F2 Recovery (2022-24)	417.3%	+183.3%	234%	—	—
F3 Lateral (2024-25)	218.6%	+218.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

32.219972142219945

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r146_atr_stops.py

Experimento: R146: ATR-adaptive initial SL on canonical V66
Origen: outputs/round_146_atr_stops/metadata.json · → ver detalle de la ronda

R150 DD 20% / pause 30dBUGGY-ENGINE 12,491% +70.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	169.8%	+65.8%	104%	—	—
F1 Bull (2019-21)	168.8%	-993.2%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	202.7%	+202.7%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

20.29053530619941

circuit_triggers

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r150_maxdd_circuit.py

Experimento: R150: Max-DD circuit breaker on V66
Origen: outputs/round_150_maxdd/metadata.json · → ver detalle de la ronda

R150 DD 15% / pause 30dBUGGY-ENGINE 6,938% -29.8 3/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	168.7%	+64.7%	104%	—	—
F1 Bull (2019-21)	68.2%	-1,093.8%	1162%	—	—
F2 Recovery (2022-24)	349.7%	+115.7%	234%	—	—
F3 Lateral (2024-25)	246.4%	+246.4%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

avg_mdd

17.41430641438786

circuit_triggers

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r150_maxdd_circuit.py

Experimento: R150: Max-DD circuit breaker on V66
Origen: outputs/round_150_maxdd/metadata.json · → ver detalle de la ronda

R162 R134 size: hard 25/50/75/100BUGGY-ENGINE 6,278% -90.4 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	65.8%	-1,096.2%	1162%	—	—
F2 Recovery (2022-24)	144.0%	-90.0%	234%	—	—
F3 Lateral (2024-25)	194.6%	+194.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r162_r134_size_multiplier.py

Experimento: R162: R134 P(v66_zone) as position size multiplier
Origen: outputs/round_162_r134_size_mult/metadata.json · → ver detalle de la ronda

R162 R134 size: soft (linear P)BUGGY-ENGINE 4,388% -116.6 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	62.8%	-1,099.2%	1162%	—	—
F2 Recovery (2022-24)	117.8%	-116.2%	234%	—	—
F3 Lateral (2024-25)	136.5%	+136.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r162_r134_size_multiplier.py

Experimento: R162: R134 P(v66_zone) as position size multiplier
Origen: outputs/round_162_r134_size_mult/metadata.json · → ver detalle de la ronda

R148 (unnamed)BUGGY-ENGINE 3,810% -111.6 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	-9.0%	-1,171.0%	1162%	—	—
F2 Recovery (2022-24)	122.8%	-111.2%	234%	—	—
F3 Lateral (2024-25)	260.4%	+260.4%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

tau

0.2

trades_total

182

skip_pct

94.4206008583691

avg_mdd

22.90006047409818

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r148_meta_label.py

Experimento: R148: meta-labeling V66 + R134 P(v66_zone) filter
Origen: outputs/round_148_metalabel/metadata.json · → ver detalle de la ronda

R148 (unnamed)BUGGY-ENGINE 3,681% -115.8 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	-17.8%	-1,179.8%	1162%	—	—
F2 Recovery (2022-24)	140.0%	-94.0%	234%	—	—
F3 Lateral (2024-25)	258.1%	+258.1%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

tau

0.1

trades_total

207

skip_pct

92.28187919463086

avg_mdd

25.836237270527242

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r148_meta_label.py

Experimento: R148: meta-labeling V66 + R134 P(v66_zone) filter
Origen: outputs/round_148_metalabel/metadata.json · → ver detalle de la ronda

R162 R134 size: aggressive 10/30/60/100BUGGY-ENGINE 3,415% -133.6 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	20.8%	-1,141.2%	1162%	—	—
F2 Recovery (2022-24)	100.8%	-133.2%	234%	—	—
F3 Lateral (2024-25)	170.8%	+170.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r162_r134_size_multiplier.py

Experimento: R162: R134 P(v66_zone) as position size multiplier
Origen: outputs/round_162_r134_size_mult/metadata.json · → ver detalle de la ronda

R148 (unnamed)BUGGY-ENGINE 2,921% -147.1 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	-12.0%	-1,174.0%	1162%	—	—
F2 Recovery (2022-24)	87.3%	-146.7%	234%	—	—
F3 Lateral (2024-25)	242.5%	+242.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

tau

0.3

trades_total

167

skip_pct

95.26241134751773

avg_mdd

23.11876349394448

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r148_meta_label.py

Experimento: R148: meta-labeling V66 + R134 P(v66_zone) filter
Origen: outputs/round_148_metalabel/metadata.json · → ver detalle de la ronda

R148 (unnamed)BUGGY-ENGINE 2,360% -158.2 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	2.7%	-1,159.3%	1162%	—	—
F2 Recovery (2022-24)	76.2%	-157.8%	234%	—	—
F3 Lateral (2024-25)	154.2%	+154.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

tau

0.7

trades_total

121

skip_pct

97.26182394206833

avg_mdd

19.861826425315833

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r148_meta_label.py

Experimento: R148: meta-labeling V66 + R134 P(v66_zone) filter
Origen: outputs/round_148_metalabel/metadata.json · → ver detalle de la ronda

R158 R151-A (canonical R069 for shorts) [baseline]BUGGY-ENGINE 1,772% -290.9 3/4 398 sh 51.2% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	483.0%	+379.0%	104%	—	—
F1 Bull (2019-21)	229.0%	-933.0%	1162%	—	—
F2 Recovery (2022-24)	-56.5%	-290.5%	234%	—	—
F3 Lateral (2024-25)	124.4%	+124.4%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=398 · WR=51.2232321592815%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=15, k_dn=5 (longer up barrier, much tighter down)

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r158_dual_r157.py

Experimento: R158: V66 canonical longs + R157 short-trained shorts
Origen: outputs/round_158_dual_model/metadata.json · → ver detalle de la ronda

R158 R157 P(SL)>0.65, exit P(TP)>0.35, sl=8%BUGGY-ENGINE 1,772% -290.9 3/4 398 sh 51.2% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	483.0%	+379.0%	104%	—	—
F1 Bull (2019-21)	229.0%	-933.0%	1162%	—	—
F2 Recovery (2022-24)	-56.5%	-290.5%	234%	—	—
F3 Lateral (2024-25)	124.4%	+124.4%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=398 · WR=51.2232321592815%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=15, k_dn=5 (longer up barrier, much tighter down)

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r158_dual_r157.py

Experimento: R158: V66 canonical longs + R157 short-trained shorts
Origen: outputs/round_158_dual_model/metadata.json · → ver detalle de la ronda

R148 (unnamed)BUGGY-ENGINE 1,576% -179.8 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	-12.0%	-1,174.0%	1162%	—	—
F2 Recovery (2022-24)	54.6%	-179.4%	234%	—	—
F3 Lateral (2024-25)	130.3%	+130.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

tau

0.4

trades_total

156

skip_pct

95.97419354838709

avg_mdd

25.50800531681281

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r148_meta_label.py

Experimento: R148: meta-labeling V66 + R134 P(v66_zone) filter
Origen: outputs/round_148_metalabel/metadata.json · → ver detalle de la ronda

R148 (unnamed)BUGGY-ENGINE 1,331% -169.8 2/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	-14.3%	-1,176.3%	1162%	—	—
F2 Recovery (2022-24)	64.6%	-169.4%	234%	—	—
F3 Lateral (2024-25)	89.6%	+89.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

tau

0.5

trades_total

146

skip_pct

96.44509374239104

avg_mdd

24.630217400457944

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r148_meta_label.py

Experimento: R148: meta-labeling V66 + R134 P(v66_zone) filter
Origen: outputs/round_148_metalabel/metadata.json · → ver detalle de la ronda

R176 R176: V66 backtest with R175 Transformer checkpoints (bug-fi 92% -223.9 0/4 243 — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	-21.2%	104%	—	—
F1 Bull (2019-21)	—	-68.9%	1162%	—	—
F2 Recovery (2022-24)	—	-223.9%	234%	—	—
F3 Lateral (2024-25)	—	-6.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=243 · WR=—

Métricas y config crudas (metadata.json)

folds_trades

[51, 136, 49, 7]

total_wins

114

v66_bugfix_baseline

{"compound": 42636, "min_alpha": 193}

beats_v66

False

elapsed_min

42.88139955202738

Reproducción (cómo replicar este modelo)

Arquitectura

Transformer encoder 4 layers × d_model=64 × 8 heads × FFN=128 (155K params). Inputs same as GRU, output 3-class. NEW arch.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam 1e-4, GELU, 5 seeds, same data + events + labeling as default.

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Transformer outputs fed into V66 cascade. FAILED: distribution mismatch (P(TP) p50=0.18 vs GRU 0.30) breaks thresholds.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r176_transformer_eval.py

Experimento: R176: V66 backtest with R175 Transformer checkpoints (bug-fixed engine)
Origen: outputs/round_176_transformer_eval/metadata.json · → ver detalle de la ronda

R158 R157 P(SL)>0.60, exit P(TP)>0.30, sl=8%BUGGY-ENGINE -12% -307.2 2/4 1,540 sh 45.8% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	58.2%	-45.8%	104%	—	—
F1 Bull (2019-21)	13.7%	-1,148.3%	1162%	—	—
F2 Recovery (2022-24)	-72.8%	-306.8%	234%	—	—
F3 Lateral (2024-25)	80.7%	+80.7%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=1540 · WR=45.845186538028486%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=15, k_dn=5 (longer up barrier, much tighter down)

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r158_dual_r157.py

Experimento: R158: V66 canonical longs + R157 short-trained shorts
Origen: outputs/round_158_dual_model/metadata.json · → ver detalle de la ronda

R169 R169: V66 backtest with R168 MTF-trained checkpointsBUGGY-ENGINE -56% -222.2 0/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	-61.7%	104%	—	—
F1 Bull (2019-21)	—	-139.7%	1162%	—	—
F2 Recovery (2022-24)	—	-222.2%	234%	—	—
F3 Lateral (2024-25)	—	-34.9%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

folds_trades

[1, 85, 41, 46]

r136_baseline

{"compound": 70576, "min_alpha": 245, "folds_alpha": [378...

improves_vs_r136

False

verdict

FAIL — MTF doesn't help

elapsed_min

7.2285001397132875

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r169_mtf_eval.py

Experimento: R169: V66 backtest with R168 MTF-trained checkpoints
Origen: outputs/round_169_mtf_eval/metadata.json · → ver detalle de la ronda

R158 R157 P(SL)>0.55, exit P(TP)>0.35, sl=8%BUGGY-ENGINE -96% -303.8 1/4 900 sh 47.5% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	-57.3%	-161.3%	104%	—	—
F1 Bull (2019-21)	-86.4%	-1,248.4%	1162%	—	—
F2 Recovery (2022-24)	-69.4%	-303.4%	234%	—	—
F3 Lateral (2024-25)	116.9%	+116.9%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=900 · WR=47.514101503072084%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=15, k_dn=5 (longer up barrier, much tighter down)

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r158_dual_r157.py

Experimento: R158: V66 canonical longs + R157 short-trained shorts
Origen: outputs/round_158_dual_model/metadata.json · → ver detalle de la ronda

R158 R157 P(SL)>0.50, exit P(TP)>0.40, sl=8%BUGGY-ENGINE -99% -305.4 0/4 667 sh 44.5% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	-76.5%	-180.5%	104%	—	—
F1 Bull (2019-21)	-83.0%	-1,245.0%	1162%	—	—
F2 Recovery (2022-24)	-71.0%	-305.0%	234%	—	—
F3 Lateral (2024-25)	-29.2%	-29.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=667 · WR=44.51303463530242%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=15, k_dn=5 (longer up barrier, much tighter down)

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r158_dual_r157.py

Experimento: R158: V66 canonical longs + R157 short-trained shorts
Origen: outputs/round_158_dual_model/metadata.json · → ver detalle de la ronda

R158 R157 P(SL)>0.50, exit P(TP)>0.40, sl=10%BUGGY-ENGINE -100% -311.6 0/4 638 sh 44.6% —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	-78.7%	-182.7%	104%	—	—
F1 Bull (2019-21)	-82.3%	-1,244.3%	1162%	—	—
F2 Recovery (2022-24)	-77.2%	-311.2%	234%	—	—
F3 Lateral (2024-25)	-42.0%	-42.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=638 · WR=44.618133638296385%

Métricas y config crudas (metadata.json)

No extra config

—

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2x128 5-seed RE-TRAINED with short-specific tb3 labeling (k_up=7, k_dn=10) — asymmetric for catching drops faster.

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=15, k_dn=5 (longer up barrier, much tighter down)

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r158_dual_r157.py

Experimento: R158: V66 canonical longs + R157 short-trained shorts
Origen: outputs/round_158_dual_model/metadata.json · → ver detalle de la ronda

R135 Production-grade bar-by-bar portfolio simulatorBUGGY-ENGINE — -227.7 —/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

round

135

method

V66 GRU 5-seed inference + uniform-exit + cooldown logic,...

completed_at

2026-05-26T13:19:04.568384+00:00

elapsed_min

1.42

v66_compound

43.51348400339998

macd_compound

-89.1549372176

portfolio_compound

-60.1537811512

fold_summary

[{"fold": 0, "fold_name": "F0_Bear", "n_v66_trades": 30, ...

portfolio_results

[{"fold": 0, "fold_name": "F0_Bear", "method": "fixed 25/...

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r135.py

Experimento: Production-grade bar-by-bar portfolio simulator
Origen: outputs/round_135/metadata.json · → ver detalle de la ronda

R159 V66 canonicalPRODOOTBUGGY-ENGINE — — 1/4 — — +11.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.554368963751494

alpha

11.43278473470329

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r159_oot_2026.py

Experimento: R159: 2026 OOT validation
Origen: outputs/round_159_oot_2026/metadata.json · → ver detalle de la ronda

R159 R151-AOOTBUGGY-ENGINE — — 1/4 2 sh — +13.8%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=2 · WR=—

Métricas y config crudas (metadata.json)

pnl

-5.227596461050816

alpha

13.759557237403968

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r159_oot_2026.py

Experimento: R159: 2026 OOT validation
Origen: outputs/round_159_oot_2026/metadata.json · → ver detalle de la ronda

R159 R151-BOOTBUGGY-ENGINE — — 1/4 — — +11.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.554368963751494

alpha

11.43278473470329

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r159_oot_2026.py

Experimento: R159: 2026 OOT validation
Origen: outputs/round_159_oot_2026/metadata.json · → ver detalle de la ronda

R159 R154OOTBUGGY-ENGINE — — 1/4 2 sh — +13.8%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=2 · WR=—

Métricas y config crudas (metadata.json)

pnl

-5.227596461050816

alpha

13.759557237403968

shorts_skipped

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r159_oot_2026.py

Experimento: R159: 2026 OOT validation
Origen: outputs/round_159_oot_2026/metadata.json · → ver detalle de la ronda

R165 V66 canonicalPRODOOTBUGGY-ENGINE — — 1/4 — — +11.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.55436896375144

alpha

11.432784734703343

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r165_oot_pyramid.py

Experimento: R165: 2026 OOT for R163 pyramid variants + comparisons
Origen: outputs/round_165_oot_pyramid/metadata.json · → ver detalle de la ronda

R165 R163 pyramidOOTBUGGY-ENGINE — — 1/4 — — +60.3%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

41.30449568804348

alpha

60.291649386498264

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r165_oot_pyramid.py

Experimento: R165: 2026 OOT for R163 pyramid variants + comparisons
Origen: outputs/round_165_oot_pyramid/metadata.json · → ver detalle de la ronda

R165 R163 +5%/+30%/1.3x/12hOOTBUGGY-ENGINE — — 1/4 — — +41.1%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

22.099480702744838

alpha

41.08663440119962

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r165_oot_pyramid.py

Experimento: R165: 2026 OOT for R163 pyramid variants + comparisons
Origen: outputs/round_165_oot_pyramid/metadata.json · → ver detalle de la ronda

R165 R163 +15%/+50%/1.5x/48hOOTBUGGY-ENGINE — — 1/4 — — +11.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.55436896375144

alpha

11.432784734703343

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r165_oot_pyramid.py

Experimento: R165: 2026 OOT for R163 pyramid variants + comparisons
Origen: outputs/round_165_oot_pyramid/metadata.json · → ver detalle de la ronda

R165 R163 +20%/+100%/2.0x/48hOOTBUGGY-ENGINE — — 1/4 — — +11.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.55436896375144

alpha

11.432784734703343

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r165_oot_pyramid.py

Experimento: R165: 2026 OOT for R163 pyramid variants + comparisons
Origen: outputs/round_165_oot_pyramid/metadata.json · → ver detalle de la ronda

R165 R151-A shortsOOTBUGGY-ENGINE — — 1/4 2 sh — +13.8%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=2 · WR=—

Métricas y config crudas (metadata.json)

pnl

-5.227596461050816

alpha

13.759557237403968

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r165_oot_pyramid.py

Experimento: R165: 2026 OOT for R163 pyramid variants + comparisons
Origen: outputs/round_165_oot_pyramid/metadata.json · → ver detalle de la ronda

R171 V66 canonicalPRODOOTBUGGY-ENGINE — — —/4 — — +11.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.55436896375144

alpha

11.432784734703343

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r171_oot_new_champ.py

Experimento: R171: OOT 2026 for R170 sweep winners
Origen: outputs/round_171_oot_new_champ/metadata.json · → ver detalle de la ronda

R171 R163 baseline (0.10/0.50/1.5x/24h)OOTBUGGY-ENGINE — — —/4 — — +60.3%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

41.30449568804348

alpha

60.291649386498264

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r171_oot_new_champ.py

Experimento: R171: OOT 2026 for R170 sweep winners
Origen: outputs/round_171_oot_new_champ/metadata.json · → ver detalle de la ronda

R171 R170-A (0.05/0.50/1.5x/24h) ★OOTBUGGY-ENGINE — — —/4 — — +60.9%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

41.90008113313115

alpha

60.887234831585936

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r171_oot_new_champ.py

Experimento: R171: OOT 2026 for R170 sweep winners
Origen: outputs/round_171_oot_new_champ/metadata.json · → ver detalle de la ronda

R171 R170-B (0.12/0.65/1.75x/24h)OOTBUGGY-ENGINE — — —/4 — — +71.9%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

52.95842635339719

alpha

71.94558005185198

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r171_oot_new_champ.py

Experimento: R171: OOT 2026 for R170 sweep winners
Origen: outputs/round_171_oot_new_champ/metadata.json · → ver detalle de la ronda

R171 R170-C (0.10/0.50/1.75x/24h)OOTBUGGY-ENGINE — — —/4 — — +60.3%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

41.30449568804348

alpha

60.291649386498264

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r171_oot_new_champ.py

Experimento: R171: OOT 2026 for R170 sweep winners
Origen: outputs/round_171_oot_new_champ/metadata.json · → ver detalle de la ronda

R171 R170-D (0.10/0.25/1.5x/24h)OOTBUGGY-ENGINE — — —/4 — — +59.1%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

40.11172246397922

alpha

59.098876162434

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r171_oot_new_champ.py

Experimento: R171: OOT 2026 for R170 sweep winners
Origen: outputs/round_171_oot_new_champ/metadata.json · → ver detalle de la ronda

R171 R170-E (0.08/0.50/1.5x/24h)OOTBUGGY-ENGINE — — —/4 — — +60.9%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

41.90794090857351

alpha

60.89509460702829

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r171_oot_new_champ.py

Experimento: R171: OOT 2026 for R170 sweep winners
Origen: outputs/round_171_oot_new_champ/metadata.json · → ver detalle de la ronda

R172 R172: R170-B with full instrumentation per foldBUGGY-ENGINE — +415.5 4/4 301 — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=301 · WR=—

Métricas y config crudas (metadata.json)

config

{"trigger": 0.12, "add_size": 0.65, "max_lev": 1.75, "min...

compound_total

257753.52085323367

simple_sum_total

759.4234047223401

total_wins

166

avg_wr_unweighted

55.42761418999043

elapsed_min

5.9215695063273115

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

(Instrumented re-run of R170-B for audit. Per-trade logs revealed phantom adds → trigger for R173.)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r172_instrumented.py

Experimento: R172: R170-B with full instrumentation per fold
Origen: outputs/round_172_instrumented/metadata.json · → ver detalle de la ronda

R177 V66 alone bugfixPRODOOT — — 1/4 10 — +12.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=10 · WR=—

Métricas y config crudas (metadata.json)

pnl

-6.620420847287624

alpha

12.366732851167159

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

OOT 2026: full hold-out (2026-01-01 to 2026-03-23). NO training on this period.

Script fuente

scripts/run_r177_oot_r151_bugfix.py

Experimento: R177: OOT 2026 R151-A under bug-fixed engine
Origen: outputs/round_177_oot_r151_bugfix/metadata.json · → ver detalle de la ronda

R177 R151-A shorts full notional (bugfix)CANDIDATEOOT — — 1/4 2 sh — +15.3%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=2 · WR=—

Métricas y config crudas (metadata.json)

pnl

-3.685996882913823

alpha

15.30115681554096

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

OOT 2026: full hold-out (2026-01-01 to 2026-03-23). NO training on this period.

Script fuente

scripts/run_r177_oot_r151_bugfix.py

Experimento: R177: OOT 2026 R151-A under bug-fixed engine
Origen: outputs/round_177_oot_r151_bugfix/metadata.json · → ver detalle de la ronda

R177 R151-A half-Kelly (bugfix)CANDIDATEOOT — — 1/4 2 sh — +11.5%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=2 · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.50420184892284

alpha

11.482951849531943

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

OOT 2026: full hold-out (2026-01-01 to 2026-03-23). NO training on this period.

Script fuente

scripts/run_r177_oot_r151_bugfix.py

Experimento: R177: OOT 2026 R151-A under bug-fixed engine
Origen: outputs/round_177_oot_r151_bugfix/metadata.json · → ver detalle de la ronda

R177 R151-B conservative (bugfix)CANDIDATEOOT — — 1/4 — — +12.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-6.620420847287624

alpha

12.366732851167159

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R151-A (V66 longs + shorts in V66 cash gaps) under bug-fixed engine. Honest re-validation of R151 family.

Walk-forward

OOT 2026: full hold-out (2026-01-01 to 2026-03-23). NO training on this period.

Script fuente

scripts/run_r177_oot_r151_bugfix.py

Experimento: R177: OOT 2026 R151-A under bug-fixed engine
Origen: outputs/round_177_oot_r151_bugfix/metadata.json · → ver detalle de la ronda

R166 COMBINED R163 + R151-APHANTOM 454,068% +287.2 4/4 93 sh — —

🚨 PHANTOM — Esta estrategia fue invalidada por la auditoría R173 (intra-bar lookahead + apalancamiento sin coste). El compound mostrado es 100% artefacto del engine bug. NO compares con números honestos. Si te interesa el enfoque, mira la versión bug-fixed (R173+) si existe.

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	1,321.5%	+1,217.5%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	753.4%	+519.4%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r166_combined.py

Experimento: R166: combined R163 pyramid + R151-A shorts (stacking test)
Origen: outputs/round_166_combined/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.50 · lev=1.75x · hold=24hPHANTOM 269,825% +372.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	675.8%	+571.8%	104%	—	—
F1 Bull (2019-21)	764.5%	-397.5%	1162%	—	—
F2 Recovery (2022-24)	740.9%	+506.9%	234%	—	—
F3 Lateral (2024-25)	378.6%	+378.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.5

max_lev

1.75

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.12 · add=0.65 · lev=1.75x · hold=24hPHANTOM 257,754% +415.5 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	748.4%	+644.4%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	814.8%	+580.8%	234%	—	—
F3 Lateral (2024-25)	421.3%	+421.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.12

add_size

0.65

max_lev

1.75

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.25 · lev=1.50x · hold=24hPHANTOM 231,575% +371.1 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	673.2%	+569.2%	104%	—	—
F1 Bull (2019-21)	650.7%	-511.3%	1162%	—	—
F2 Recovery (2022-24)	736.9%	+502.9%	234%	—	—
F3 Lateral (2024-25)	376.9%	+376.9%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.25

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.05 · add=0.50 · lev=1.50x · hold=24hPHANTOM 228,349% +430.4 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	685.2%	+581.2%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	751.4%	+517.4%	234%	—	—
F3 Lateral (2024-25)	436.2%	+436.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.05

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.08 · add=0.40 · lev=1.50x · hold=24hPHANTOM 226,896% +346.6 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	635.1%	+531.1%	104%	—	—
F1 Bull (2019-21)	761.8%	-400.2%	1162%	—	—
F2 Recovery (2022-24)	692.1%	+458.1%	234%	—	—
F3 Lateral (2024-25)	352.4%	+352.4%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.08

add_size

0.4

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R163 pyramid +20% / +100% / 2.0x / 48hPHANTOM 219,361% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	898.0%	+794.0%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	501.5%	+501.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

total_pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 + PYRAMIDING longs: cuando price > entry*(1+trigger) (e.g., +10%), add size_frac of remaining cap; up to max_leverage. ⚠️ INTRA-BAR LOOKAHEAD BUG in entry sequencing — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r163_pyramiding.py

Experimento: R163: Pyramiding × always-invested V66
Origen: outputs/round_163_pyramiding/metadata.json · → ver detalle de la ronda

R166 R151-A shorts only (P(SL)>0.65)PHANTOM 212,591% +245.1 4/4 93 sh — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	880.5%	+776.5%	104%	—	—
F1 Bull (2019-21)	852.5%	-309.5%	1162%	—	—
F2 Recovery (2022-24)	479.5%	+245.5%	234%	—	—
F3 Lateral (2024-25)	293.0%	+293.0%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=93 · WR=—

Métricas y config crudas (metadata.json)

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r166_combined.py

Experimento: R166: combined R163 pyramid + R151-A shorts (stacking test)
Origen: outputs/round_166_combined/metadata.json · → ver detalle de la ronda

R170 trig=0.08 · add=0.50 · lev=1.50x · hold=24hPHANTOM 204,690% +377.5 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	685.2%	+581.2%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	746.7%	+512.7%	234%	—	—
F3 Lateral (2024-25)	383.3%	+383.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.08

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.08 · add=0.50 · lev=1.50x · hold=24hPHANTOM 204,690% +377.5 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	685.2%	+581.2%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	746.7%	+512.7%	234%	—	—
F3 Lateral (2024-25)	383.3%	+383.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.08

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R163 pyramid +10% / +50% / 1.5x / 24hPHANTOM 198,893% +372.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	675.8%	+571.8%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	740.9%	+506.9%	234%	—	—
F3 Lateral (2024-25)	378.6%	+378.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

total_pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 + PYRAMIDING longs: cuando price > entry*(1+trigger) (e.g., +10%), add size_frac of remaining cap; up to max_leverage. ⚠️ INTRA-BAR LOOKAHEAD BUG in entry sequencing — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r163_pyramiding.py

Experimento: R163: Pyramiding × always-invested V66
Origen: outputs/round_163_pyramiding/metadata.json · → ver detalle de la ronda

R166 R163 pyramid only (+10/+50/1.5x/24h)PHANTOM 198,893% +372.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	675.8%	+571.8%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	740.9%	+506.9%	234%	—	—
F3 Lateral (2024-25)	378.6%	+378.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r166_combined.py

Experimento: R166: combined R163 pyramid + R151-A shorts (stacking test)
Origen: outputs/round_166_combined/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.50 · lev=1.50x · hold=24hPHANTOM 198,893% +372.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	675.8%	+571.8%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	740.9%	+506.9%	234%	—	—
F3 Lateral (2024-25)	378.6%	+378.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.50 · lev=1.50x · hold=12hPHANTOM 198,893% +372.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	675.8%	+571.8%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	740.9%	+506.9%	234%	—	—
F3 Lateral (2024-25)	378.6%	+378.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.50 · lev=1.50x · hold=48hPHANTOM 198,893% +372.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	675.8%	+571.8%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	740.9%	+506.9%	234%	—	—
F3 Lateral (2024-25)	378.6%	+378.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.12 · add=0.50 · lev=1.50x · hold=24hPHANTOM 196,935% +371.0 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	675.8%	+571.8%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	735.9%	+501.9%	234%	—	—
F3 Lateral (2024-25)	376.8%	+376.8%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.12

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R163 pyramid +15% / +50% / 1.5x / 48hPHANTOM 191,642% +366.3 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	670.4%	+566.4%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	727.3%	+493.3%	234%	—	—
F3 Lateral (2024-25)	372.1%	+372.1%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

total_pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 + PYRAMIDING longs: cuando price > entry*(1+trigger) (e.g., +10%), add size_frac of remaining cap; up to max_leverage. ⚠️ INTRA-BAR LOOKAHEAD BUG in entry sequencing — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r163_pyramiding.py

Experimento: R163: Pyramiding × always-invested V66
Origen: outputs/round_163_pyramiding/metadata.json · → ver detalle de la ronda

R170 trig=0.15 · add=0.50 · lev=1.50x · hold=24hPHANTOM 191,642% +366.3 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	670.4%	+566.4%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	727.3%	+493.3%	234%	—	—
F3 Lateral (2024-25)	372.1%	+372.1%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.15

add_size

0.5

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.40 · lev=1.50x · hold=24hPHANTOM 163,654% +342.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	627.5%	+523.5%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	687.3%	+453.3%	234%	—	—
F3 Lateral (2024-25)	348.6%	+348.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.4

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.15 · add=0.40 · lev=1.50x · hold=48hPHANTOM 158,495% +337.5 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	623.2%	+519.2%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	676.3%	+442.3%	234%	—	—
F3 Lateral (2024-25)	343.3%	+343.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.15

add_size

0.4

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R163 pyramid +5% / +30% / 1.3x / 12hPHANTOM 150,563% +360.4 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	585.0%	+481.0%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	640.3%	+406.3%	234%	—	—
F3 Lateral (2024-25)	366.2%	+366.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

total_pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 + PYRAMIDING longs: cuando price > entry*(1+trigger) (e.g., +10%), add size_frac of remaining cap; up to max_leverage. ⚠️ INTRA-BAR LOOKAHEAD BUG in entry sequencing — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r163_pyramiding.py

Experimento: R163: Pyramiding × always-invested V66
Origen: outputs/round_163_pyramiding/metadata.json · → ver detalle de la ronda

R170 trig=0.05 · add=0.30 · lev=1.30x · hold=12hPHANTOM 150,563% +360.4 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	585.0%	+481.0%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	640.3%	+406.3%	234%	—	—
F3 Lateral (2024-25)	366.2%	+366.2%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.05

add_size

0.3

max_lev

1.3

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.15 · add=0.30 · lev=1.30x · hold=48hPHANTOM 129,452% +308.7 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	576.0%	+472.0%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	625.5%	+391.5%	234%	—	—
F3 Lateral (2024-25)	314.5%	+314.5%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.15

add_size

0.3

max_lev

1.3

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R163 pyramid +10% / +25% / 1.25x / 24hPHANTOM 119,050% +297.8 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	555.2%	+451.2%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	607.1%	+373.1%	234%	—	—
F3 Lateral (2024-25)	303.6%	+303.6%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

total_pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 + PYRAMIDING longs: cuando price > entry*(1+trigger) (e.g., +10%), add size_frac of remaining cap; up to max_leverage. ⚠️ INTRA-BAR LOOKAHEAD BUG in entry sequencing — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r163_pyramiding.py

Experimento: R163: Pyramiding × always-invested V66
Origen: outputs/round_163_pyramiding/metadata.json · → ver detalle de la ronda

R163 canonical V66 (no pyramid)PHANTOM 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

total_pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

V66 + PYRAMIDING longs: cuando price > entry*(1+trigger) (e.g., +10%), add size_frac of remaining cap; up to max_leverage. ⚠️ INTRA-BAR LOOKAHEAD BUG in entry sequencing — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r163_pyramiding.py

Experimento: R163: Pyramiding × always-invested V66
Origen: outputs/round_163_pyramiding/metadata.json · → ver detalle de la ronda

R166 canonical V66PHANTOM 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r166_combined.py

Experimento: R166: combined R163 pyramid + R151-A shorts (stacking test)
Origen: outputs/round_166_combined/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.65 · lev=1.50x · hold=24hPHANTOM 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.65

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.80 · lev=1.50x · hold=24hPHANTOM 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.8

max_lev

1.5

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.50 · lev=1.25x · hold=24hPHANTOM 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.5

max_lev

1.25

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.05 · add=0.40 · lev=1.30x · hold=12hPHANTOM 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.05

add_size

0.4

max_lev

1.3

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R170 trig=0.10 · add=0.80 · lev=1.75x · hold=24hPHANTOM 70,576% +239.2 4/4 — — —

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	435.1%	+331.1%	104%	—	—
F1 Bull (2019-21)	537.3%	-624.7%	1162%	—	—
F2 Recovery (2022-24)	473.6%	+239.6%	234%	—	—
F3 Lateral (2024-25)	261.3%	+261.3%	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

trigger

0.1

add_size

0.8

max_lev

1.75

min_hold

total_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

Dense sweep over R163 pyramid params (trigger/add_size/max_lev/min_hold). ⚠️ Same intra-bar bug. ALL 20 variants INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r170_pyramid_sweep.py

Experimento: R170: dense pyramid sweep around R163 winner
Origen: outputs/round_170_pyramid_sweep/metadata.json · → ver detalle de la ronda

R167 V66 canonicalPHANTOMOOT — — 1/4 — — +11.4%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

-7.55436896375144

alpha

11.432784734703343

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r167_oot_combined.py

Experimento: R167: OOT 2026 for COMBINED + comparison
Origen: outputs/round_167_oot_combined/metadata.json · → ver detalle de la ronda

R167 R163 pyramid onlyPHANTOMOOT — — 1/4 — — +60.3%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=— · WR=—

Métricas y config crudas (metadata.json)

pnl

41.30449568804348

alpha

60.291649386498264

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r167_oot_combined.py

Experimento: R167: OOT 2026 for COMBINED + comparison
Origen: outputs/round_167_oot_combined/metadata.json · → ver detalle de la ronda

R167 R151-A shorts onlyPHANTOMOOT — — 1/4 2 sh — +13.8%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=2 · WR=—

Métricas y config crudas (metadata.json)

pnl

-5.2275964610507435

alpha

13.759557237404039

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r167_oot_combined.py

Experimento: R167: OOT 2026 for COMBINED + comparison
Origen: outputs/round_167_oot_combined/metadata.json · → ver detalle de la ronda

R167 COMBINED R163+R151-APHANTOMOOT — — 1/4 2 sh — +13.8%

Performance por fold

Fold	PnL	α	BH	Trades	WR
F0 Bear (2018)	—	—	104%	—	—
F1 Bull (2019-21)	—	—	1162%	—	—
F2 Recovery (2022-24)	—	—	234%	—	—
F3 Lateral (2024-25)	—	—	0%	—	—

Trades/WR por fold solo disponibles cuando metadata.json los registra (R180+). Totales: trades=2 · WR=—

Métricas y config crudas (metadata.json)

pnl

-5.2275964610507435

alpha

13.759557237404039

pyramid_adds

Reproducción (cómo replicar este modelo)

Arquitectura

GRU 2 layers × 128 hidden × 5-seed ensemble (checkpoints en outputs/round_069)

Indicadores / features

RSI(14), MACD, Bollinger %B, ATR_14_pct (vol normalization), daily SMA20+pct_change5 slope, DRSI(28) daily, avg_tp_bar, plus 30+ rolling stats from src/data/features.py::add_all_features()

Entrenamiento

Adam lr=1e-3, 50 epochs early-stop on val_loss, batch=64, dropout=0.2, 5-seed ensemble (seeds 17/29/41/53/67)

Filtro de eventos

Volatility event filter: atr_mult=2.0 + return_percentile=90 → ~10.5% of bars selected

Labeling

Triple-barrier k_up=10 × k_dn=7 (multiples of ATR_14_pct), 3-class output

Entrada / salida

R163 pyramid + R151-A shorts combined. ⚠️ Combined two bugs — INVALIDATED.

Walk-forward

Walk-forward 4 folds, val_pct=10%, no purging beyond chronological cut

Script fuente

scripts/run_r167_oot_combined.py

Experimento: R167: OOT 2026 for COMBINED + comparison
Origen: outputs/round_167_oot_combined/metadata.json · → ver detalle de la ronda