CURRENT CHAMPION

V6 · V66 Cooldown(4,48)

+70,576% compound · +245% min α · 4/4 beats B&H · MDD 27.7%

GRU 2x128 ensemble · 5 seeds × 4 folds · 36 features · tb_vol_10_7 labeling

Last round R119 MTF filter · REJECT

Rounds since V66 beaten 8 paradigm exhaustion

Active bots 5/5 all running

Live capital €125 5 × €25 sub-accounts

R&D signal: 8 rounds without beating V66 baseline. Refinement is exhausted — shift focus to NEW paradigms. R117 MACD_div (+179K% compound, decorrelated from V66) is the live lead — consider R120 as portfolio combination V66 + MACD_div.

Production bots 5 on Hetzner CX23 · all share GRU checkpoints from R069

GRU Ensemble (always-invested)

compound+7,641% min α— tf15min configproduction_v2.yaml

GRU + DailyRSI>80 + Trail 10%

compound+450B% min α— tf15min configproduction_v3_hybrid.yam

GRU 3-Regime adaptive (V5.4 Robust-5)

compound+30,174% min α+117% tf15min configproduction_v4_hybrid_v5.

V115_cmp (GRU + peak_drop + ratchet + regime cooldown)

compound**+168,759%** min α**+238%** tf15min configproduction_v5_hybrid_v11

V66 Cooldown(4,48) (GRU defensive uniform)

compound+70,576% min α**+245%** tf15min configproduction_v6_hybrid_v66

Latest experimental results most recent 5 rounds with full metrics

R119 mtf-filter SUPERSEDED

2026-05-21

Multi-timeframe filter (4h context on V66 15min)

filter	compound_pnl	min_alpha	fold_0_alpha	fold_1_alpha	fold_2_alpha	fold_3_alpha	beats_bh_folds	total_trades	blocked_entries	avg_wr	avg_mdd	time_s
baseline_V66	70576.0	245.2	375.9	437.3	245.2	255.5	4	300	0	55.9	27.7	1.9
MTF_EMA_bullish	6730.1	-52.3	223.1	196.5	-52.3	57.3	3	229	3057	53.0	27.8	2.3
MTF_no_extreme_RSI	66749.6	168.8	354.4	597.1	168.8	222.6	4	286	722	55.2	26.1	3.5
MTF_slope_supportive	12583.7	72.9	227.3	248.9	80.7	72.9	4	243	2137	54.5	24.4	1.9
MTF_vol_regime	31539.2	161.3	186.4	405.2	238.1	161.3	4	286	807	55.5	26.7	1.9
MTF_all	1379.2	-119.8	49.5	39.6	-119.8	35.9	3	191	4570	46.9	25.1	1.8

R118 alt-model REJECT

2026-05-21

XGBoost classifier as GRU alternative (EXPLORATION new model family)

fold	pnl	trades	wr	mdd	BH	alpha	baseline_pnl	baseline_alpha	pnl_delta_vs_baseline
0	-1.031730088176082	248	52.41935483870967	45.84109021564403	59.18068329501152	-60.2124133831876	435.0806832950115	375.9	-436.1124133831876
1	73.6333527544727	80	60.0	8.41744916433773	99.98283255066114	-26.349479796188447	537.2828325506612	437.3	-463.6494797961885
2	100.25360241862897	384	47.13541666666667	32.089134778691964	228.4183620424153	-128.16475962378632	473.6183620424153	245.2	-373.3647596237863
3	0.49494430155902597	4	75.0	2.22385590884031	5.8144742341911915	-5.3195299326321654	261.3144742341912	255.5	-260.8195299326322

R117 mean-reversion ITERATE

2026-05-21

Mean reversion strategies portfolio (EXPLORATION — new paradigm)

strategy	compound_pnl	min_alpha	beats_bh_folds	total_trades	avg_wr	avg_mdd	fold_0_pnl	fold_1_pnl	fold_2_pnl	fold_3_pnl
BB_revert	-100.0	-327.5	0	3342	42.9	99.0	-99.0	-99.0	-99.0	-99.0
RSI_extreme	-100.0	-319.1	0	1504	45.1	94.7	-98.5	-96.6	-90.7	-92.7
Z_score	-100.0	-324.6	0	2359	52.2	97.8	-99.0	-98.6	-96.2	-97.4
MACD_div	179453.2	-42.0	3	6544	44.3	18.8	539.7	2505.6	186.4	276.1
baseline_V66_reference	70576.0	245.2	4	300	55.9	27.7	ref	ref	ref	ref

R116 sizing SUPERSEDED

2026-05-21

Position sizing (fractional Kelly, vol_scaled, dd_throttled)

config	mode	fixed_frac	compound_pnl	min_alpha	fold_0_alpha	fold_1_alpha	fold_2_alpha	fold_3_alpha	beats_bh_folds	total_trades	avg_wr	avg_mdd	time_s
baseline_fixed_1.0	fixed	1.0	70576.0	245.2	375.9	437.3	245.2	255.5	4	300	55.9	100.0	1.9
fixed_0.9	fixed	0.9	39864.7	160.7	305.3	346.0	160.7	216.4	4	300	55.9	100.0	2.9
fixed_0.8	fixed	0.8	22099.2	87.4	241.8	264.7	87.4	180.7	4	300	55.9	100.0	3.1
fixed_0.7	fixed	0.7	12008.0	23.9	185.2	192.9	23.9	148.2	4	300	55.9	100.0	1.9
fixed_0.6	fixed	0.6	6381.7	-30.9	134.8	130.0	-30.9	118.7	3	300	55.9	100.0	1.9
fixed_0.5	fixed	0.5	3304.0	-78.0	90.4	75.3	-78.0	92.0	3	300	55.9	100.0	1.9
half_kelly_dynamic	half_kelly_dynamic	0.5	3305.4	-49.0	32.3	150.5	-49.0	75.8	3	300	55.9	100.0	1.9
vol_scaled	vol_scaled	0.0	24465.1	52.1	278.7	330.4	52.1	172.2	4	300	55.9	100.0	1.9
dd_throttled	dd_throttled	0.0	23828.5	127.0	202.7	287.9	127.0	191.7	4	300	55.9	100.0	3.5

R115 other SUPERSEDED

2026-05-21

Vote threshold sweep (asymmetric)

mv_danger	mv_safe	compound_pnl	min_alpha	fold_0_alpha	fold_1_alpha	fold_2_alpha	fold_3_alpha	beats_bh_folds	total_trades	avg_wr	avg_mdd	time_s
1	1	927.7	-111.6	164.1	-89.8	-111.6	27.2	2	1102	48.1	41.9	2.3
1	2	3930.0	-36.8	252.4	0.7	-36.8	61.5	3	754	48.4	36.2	3.8
1	3	9695.2	-34.9	224.2	244.4	-34.9	90.1	3	531	49.2	29.2	2.7
1	4	8500.0	-95.1	340.2	209.4	-95.1	74.5	3	376	51.1	24.7	2.1
1	5	1013.1	-146.7	-24.1	135.1	-146.7	29.5	2	240	46.9	27.8	2.0
2	1	10345.6	3.6	551.8	164.4	3.6	15.6	4	598	48.9	39.5	2.2
2	2	11758.6	47.3	441.5	121.6	47.3	57.6	4	437	50.1	36.2	2.2
2	3	43464.9	82.5	357.1	494.8	82.5	189.7	4	336	54.1	27.9	3.8
2	4	25129.3	-58.2	621.3	323.0	-58.2	123.0	3	283	53.6	25.8	2.9
2	5	2189.0	-121.0	-14.4	186.0	-121.0	91.7	2	204	51.5	26.6	2.0
3	1	9083.4	36.1	375.0	115.5	55.6	36.1	4	426	48.2	42.3	2.2
3	2	12167.5	72.3	315.6	86.5	178.0	72.3	4	351	50.2	38.5	2.2
3	3	70576.0	245.2	375.9	437.3	245.2	255.5	4	300	55.9	27.7	2.1
3	4	32357.8	34.1	645.0	278.6	34.1	126.8	4	264	53.5	26.3	3.3
3	5	2059.9	-104.8	-9.1	153.1	-104.8	76.5	2	202	51.2	28.7	2.9
4	1	5564.4	19.2	222.0	103.8	62.7	19.2	4	384	48.3	44.8	2.2
4	2	10806.3	73.1	288.4	75.9	165.3	73.1	4	340	50.1	38.9	2.2
4	3	57769.0	182.1	396.4	423.9	251.4	182.1	4	296	55.4	27.7	2.1
4	4	33377.4	35.4	694.1	278.6	35.4	119.6	4	263	54.3	25.8	2.0
4	5	1977.7	-101.0	-12.6	153.1	-101.0	70.6	2	202	50.5	28.9	2.9
5	1	2843.1	0.4	77.2	103.8	57.5	0.4	4	372	46.1	48.8	3.5
5	2	7746.9	51.0	201.4	75.9	174.4	51.0	4	334	50.1	41.6	2.2
5	3	50830.5	171.1	327.0	423.9	278.0	171.1	4	292	55.5	29.6	2.2
5	4	32276.2	32.7	673.2	278.6	32.7	119.3	4	261	54.4	25.9	2.1
5	5	2081.1	-101.4	-5.0	153.1	-101.4	70.6	2	200	49.7	28.9	2.1

R&D rounds 40 executed · 3 rejected · 4 records

All Record Active Iterate Reject Superseded Parked Pending

Round	Date	Paradigm	Title	Best compound	Min α	Folds BH	Avg MDD	Verdict
`R000`	2026-03-27	sizing	Baseline Setup + Binary Labeling	—	—	—	—	REJECT
`R001`	2026-03-28	hybrid-sma	Triple Barrier Labeling	—	—	—	—	—
`R002`	2026-03-28	labeling	Class Weights + Long Only + Anti-Overfit	—	—	—	—	—
`R003`	2026-03-28	labeling	GRU Simplificado + Weight Decay	—	—	—	—	PENDING
`R004`		mtf-filter	Volatility-Adaptive Triple Barrier (tb_vol)	—	—	—	—	—
`R027`		other	Post-fix: Swing, Dual-Head, Regression	—	—	—	—	—
`R036`	2026-04-02	architecture	MFE/MAE Dual-Head Regression (nuevo paradigma)	—	—	—	—	—
`R037`	2026-04-03	architecture	Walk-Forward Validation (MFE/MAE + mse_ratio)	—	—	—	—	—
`R038`	2026-04-03	other	Loss Function Comparison	—	—	—	—	REJECT
`R039`		labeling	Event-Driven Sequences + Walk-Forward Optimization	—	—	—	—	—
`R054`	2026-04-06	features	Normalization Study	—	—	—	—	RECORD
`R055`	2026-04-07	labeling	3-Class Labels	—	—	—	—	—
`R056`	2026-04-07	features	Feature Importance Study	—	—	—	—	—
`R057`	2026-04-07	features	Optimized Feature Sets	—	—	—	—	—
`R058`	2026-04-08	other	Compound Metric Study	—	—	—	—	—
`R059`	2026-04-08	labeling	3-Class + Clean Features Combined	—	—	—	—	—
`R060`	2026-04-08	features	Event Filter Comparison	—	—	—	—	—
`R061`	2026-04-09	exit-logic	Exit Mode Comparison	—	—	—	—	—
`R062`	2026-04-09	mtf-filter	Multi-Timeframe Features	—	—	—	—	—
`R063`	2026-04-09	exit-logic	Long/Short + SL Signal Exit	—	—	—	—	—
`R064`	2026-04-09	other	Sliding vs Expanding Window	—	—	—	—	—
`R065`	2026-04-09-10	other	Timeframe Comparison	—	—	—	—	—
`R066`	2026-04-10	sizing	Position Sizing	—	—	—	—	—
`R067`	2026-04-10	other	Multi-Confirmation Entry	—	—	—	—	—
`R068`	2026-04-10	other	Always-Invested Strategy	—	—	—	—	RECORD
`R069`	2026-04-10	ensemble	Multi-Seed Validation	—	—	—	—	—
`R070`	2026-04-10	ensemble	Ensemble of 5 Seeds	—	—	—	—	—
`R071`	2026-04-11	ensemble	Hybrid Always-Invested + Ensemble Voting	—	—	—	—	RECORD
`R072`	2026-04-11	other	Multi-Seed Validation of R071	—	—	—	—	—
`R073`		labeling	Labeling Innovation (Exhaustive)	—	—	—	—	—
`R081`	2026-04-12	other	Ultimate Production Config	—	—	—	—	RECORD
`R082`	2026-04-12	exit-logic	Fine Grid + DD Protection (PENDING)	—	—	—	—	PARKED
`R112`	2026-05-21	exit-logic	V66 cd_loss_uniform sweep	70,576%	245.2%	4	27.7%	SUPERSEDED
`R113`	2026-05-21	hybrid-sma	V66 multi-attack sweep (A1+A2+A5)	70,576%	245.2%	4	27.7%	SUPERSEDED
`R114`	2026-05-21	ensemble	Confidence-weighted voting (Agent C #1)	70,576%	245.2%	4	27.7%	SUPERSEDED
`R115`	2026-05-21	other	Vote threshold sweep (asymmetric)	70,576%	245.2%	4	27.7%	SUPERSEDED
`R116`	2026-05-21	sizing	Position sizing (fractional Kelly, vol_scaled, dd_throttled)	70,576%	245.2%	4	100.0%	SUPERSEDED
`R117`	2026-05-21	mean-reversion	Mean reversion strategies portfolio (EXPLORATION — new paradigm)	179,453%	-42.0%	3	18.8%	ITERATE
`R118`	2026-05-21	alt-model	XGBoost classifier as GRU alternative (EXPLORATION new model family)	100.3%	-128.2%	—	32.1%	REJECT
`R119`	2026-05-21	mtf-filter	Multi-timeframe filter (4h context on V66 15min)	66,750%	168.8%	4	26.1%	SUPERSEDED

Paradigm exploration explored vs unexplored families

other 10 rounds

—

labeling 6 rounds

tb_vol_10_7 (production)

features 4 rounds

clean 36 (production)

exit-logic 4 rounds

V66 Cd(4,48) · min α +245

ensemble 4 rounds

5-seed voting ≥3/5 (production)

sizing 3 rounds

100% per trade (R116 confirmed)

mtf-filter 3 rounds

MTF_no_extreme_RSI +66K% / +168 α (R119)

hybrid-sma 2 rounds

GRU+DRSI+trail · prod V3

architecture 2 rounds

GRU 2×128 (production)

mean-reversion 1 rounds

MACD_div +179K% / -42 α (R117)

alt-model 1 rounds

XGBoost +245% / -128 α (R118, REJECT)

Reinforcement learning 0 rounds

UNEXPLORED — R&D opportunity

Sentiment / LLM features 0 rounds

UNEXPLORED — R&D opportunity

On-chain features 0 rounds

UNEXPLORED — R&D opportunity

Multi-asset / pairs 0 rounds

UNEXPLORED — R&D opportunity

Transformer / TFT architecture 0 rounds

UNEXPLORED — R&D opportunity

R&D backlog next experiments — bias toward unexplored paradigms

Round	Title	Description	Priority
`R120`	Portfolio V66 + MACD_div	Test if MACD_div (R117) decorrelated returns can combine with V66 momentum for +α	high
`R121`	Phase 0 validation infrastructure	Build src/validation/ with the 5-step gate (calibration, 4-fold, multi-seed, sub-period, OOS)	high
`R122`	MACD_div fold-1 deep dive	Why does MACD_div underperform B&H in fold 1 (bull explosivo)? Add regime filter?	medium
`R123`	RL agent (PPO/DQN)	Stable-baselines3 on 15min OHLCV. Unexplored paradigm.	research
`R124`	Sentiment / LLM features	Use small LLM (Mistral 7B) on news headlines to generate sentiment features.	research
`R125`	On-chain features	Glassnode / mempool data as auxiliary inputs (whale flows, exchange balances).	research
`R126`	Multi-asset / pairs	BTC + ETH + SOL coordinated entries. Diversification or noise?	research
`R127`	Transformer / TFT	Temporal Fusion Transformer as GRU replacement. Last-tier exploration.	research