R&D Rounds

Binary 'up/down in N bars' labels are essentially noise in 5min crypto — the signal-to-noise ratio is too low for sequence models to learn anything but the prior. Pure binary classification on raw direction is a dead end without a profitability-aware label (triple barrier). Always sanity-check label class balance and overfit speed in the first 2 epochs.

Triple-barrier labels are only valid for the side they were generated for — you cannot symmetrically short on (pred<0.5). Class imbalance (62/38) plus no class weights collapses predictions to the prior. Sharpe computed on bar-by-bar equity with many zero-return bars is meaningless.

Class weights alone cannot rescue a model that has no discriminative signal — they just push the output distribution toward 0.5 without adding edge. WR going up while PF goes down is a classic 'more trades, smaller wins, bigger losses' trap. Regularization must be combined with a richer signal, not used as a substitute for one.

Shrinking the model does not save you from a label problem. GRU was a promising direction (cheaper, fewer gates) and would become the eventual workhorse, but only after the labels were upgraded to volatility-adaptive triple barriers. Architecture-only fixes cannot rescue weak labels.

Volatility-adaptive barriers are the right idea and stayed in every later paradigm — fixed barriers were never revived. GRU > LSTM was confirmed here as a permanent finding. The most important lesson is methodological: dropping timeout samples from training created gaps that the model couldn't reproduce live, inflating backtests ~60x. Always reconstruct the exact bar sequence the model will see in production.

Two principles emerged that stayed true: (1) very low trade count + very high WR is not a strategy — compounding needs frequency; (2) dual-head outputs that decompose 'how much up' vs 'how much down' carry more information than a single scalar and would later resurface as MFE/MAE. Always report trade count alongside WR/PF.

MFE/MAE is the most informative continuous target tried in the project, and the principle 'model that knows when not to trade' (drawdown protection through abstention) becomes a core later thesis. Decomposing P&L into upside and downside potential is a transferable idea worth keeping in any future labeling system.

First proof that walk-forward 4-fold robustness is achievable but tells you something important: a strategy that is profitable in every regime can still underperform B&H during mega-bulls because it is selectively long. From here on, 4/4 walk-forward becomes the minimum acceptance bar, and 'beats B&H compound' becomes the second.

Loss design matters more than architecture for regression labeling — the model will solve whatever you ask, including degenerate solutions like 'predict MAE=0 always'. When a regression target collapses, the answer is a ratio-based or contrastive loss, not more capacity. This pattern (loss-shapes-behavior) repeats with CE class weights in R055.

Event filtering is the single biggest non-label improvement in the whole project — keeping only ~10% of bars (the ones where 'something is happening') gave a >40x compound jump and remains in every production config. Principle: feed the model situations where decisions matter, not the long quiet stretches between them.

Per-sequence MinMax (renormalize inside each lookback window) prevents leakage of future statistics and is robust across regimes. Global / train-only z-score normalization, the default in many ML projects, is actively harmful here because BTC price scale changes by orders of magnitude across years. Always normalize inside the input window for non-stationary series.

The 3-class formulation is the single most important labeling decision in the project. Treating 'timeout' as its own class (rather than dropping it or folding it into binary) preserves a huge amount of information about indecisive market states. Weighted CE with up-weighted edge classes (2.0 / 0.5 / 2.0) keeps gradients focused on rare, decision-relevant outcomes.

Negative-importance features actively hurt: they let the model overfit to spurious correlations. Surprisingly, MACD and several volatility horizons were harmful — likely because they encode information already present in cleaner form elsewhere. Always validate features one-by-one; never trust the 'more is better' heuristic in noisy financial data.

There is a trade-off between feature count and trade quality vs robustness: smaller sets give higher per-trade quality but worse coverage. The optimal point depends on whether the strategy is selective or always-invested; clean 36 won because it generalized across both modes.

Compound returns multiply over folds, so a single bad fold tanks the whole number — making seed variance dangerous. Single-seed compound results are not trustworthy; you need ensembles or multi-seed validation to estimate the true distribution. This finding directly motivated R069/R070 ensemble work.

Feature set rankings depend on the label — what was 'highest quality' with binary labels (positive_only 24) loses to the broader clean 36 once labels become richer (3-class). Never freeze a feature ranking without locking the label first; re-run feature studies after major label changes.

vol_tight is the right event filter and lb=300 is the right lookback — this is the canonical recipe. The k/thr sweet spot (25/15, 0.40) was retained as a default; larger k values produce wider barriers that compound better when you also have an exit mechanism.

Adaptive exits look smart but in walk-forward they overfit per-trade noise and underperform fixed barriers calibrated to volatility. The cleaner the exit logic, the less surface area for overfitting. This is a powerful 'simpler beats complex' result that recurs throughout the project (and only gets reversed much later by V115 with its layered cascade — and only because the layers are independently validated).

When the lookback is already 300 event bars (~weeks of context), additional higher-timeframe features are redundant. The model is already seeing the trend. This negative result repeats in R119 (multi-timeframe FILTER) and is now a project-wide prior: don't stack timeframes on a model that already has a long lookback.

P(SL) is not symmetric with P(TP) — the model can predict 'this is bad' but not 'this will fall by X%'. Shorts have been tested multiple times and consistently fail; the recommendation is to treat BTC strategies as long-only until a dedicated short-only model is trained. Signal-based early exits also lose to staying invested.

More data > more recent data, in this regime — the model benefits from seeing the full BTC history including 2017-2018 cycles. The 'concept drift' argument for sliding windows did not materialize. Expanding window is the safer default.

Two permanent settings nailed down here: 15min base timeframe and the sigma clip [0.0005, 0.005]. The sigma clip is critical and easy to forget — extreme volatility values create huge labels that mess up the loss; tiny values create degenerate barriers. Always clip volatility inputs to a sane range before label generation.

Maximum compounding requires maximum exposure when the model's edge is real; fractional sizing trades long-run growth for lower DD. Pyramiding adds complexity (state management, partial fills) without proportional return. Use full position unless DD is a hard constraint — then prefer fewer better trades, not smaller ones.

Cheap, robust filters at the decision boundary (require 3-4 consecutive 'go' votes) are far more effective than complex model changes. The same principle reappears in R070 (ensemble voting threshold). When a model is noisy near the decision boundary, requiring persistence in agreement is a high-leverage fix.

This is the single biggest strategic shift in the project. In a long-run upward market like BTC, the cost of being out is much larger than the cost of being in during draw-downs. Use the model for DEFENSE (when to exit) rather than offense (when to enter). All later production configs (V2-V6) inherit this stance.

Single-seed metrics are essentially noise for this problem. The minimum acceptable validation is multi-seed (5+) per config. Even better is multi-seed multi-set (R084). This becomes a project-wide standard and the source of the R069 checkpoints that V2-V6 all share.

Ensembling pays for itself in robustness even when no single model is dramatically better. Hard voting (>=4/5) outperforms soft averaging because it filters out individual model overconfidence near the decision boundary. The 'require N agreements' lesson from R067 generalizes to ensemble members.

The combination of always-invested + ensemble + voting + wide barriers (k_up=20) is the canonical winning recipe. Fold 2 was historically the hardest to beat (2022-2024 recovery); cracking it 4/4 means the strategy works across regimes. Wide barriers + ensemble voting reduce trade frequency but each trade is high-conviction.

Even after ensembling, the seed-set choice can flip a fold. The right validation requires multiple SETS of ensembles (3 sets of 5 seeds), not just a single ensemble. This is the strictest robustness check used in the project and revealed that R071 was nearly-but-not-quite robust enough.

tb3_vol with 3 classes is a local optimum that is extremely hard to beat — and we now know why: (1) regression labels fail with GRU because CE concentrates gradients better; (2) timeouts (24% of labels) are essential information, not noise — methods that drop timeouts always underperform. The labeling search is closed; future gains have to come from features, exits, ensemble, or regime.

Stacking confirmed components multiplicatively (each ~+20-50%) without negative interactions is rare in ML — it works here because each component addresses a different failure mode (signal, features, sequence, strategy, robustness). The recipe is composable and survives stacking.

Always recompute equity by simulating bar-by-bar including the closing transaction explicitly — do not add 'position_value' at the end. The R082 bug taught the team to never trust a backtest until the close-out logic has been audited line by line. Every later 'record' (R084, R110, R120) is recomputed with the fixed equity logic.

Cross-team replication at the dollar is the gold-standard validation — when two independent code paths give the same numbers on the same checkpoints, the result is real. V66's +245% min alpha is the strongest worst-fold result in the project; V115 is higher compound but more complex. The lesson: a cascaded exit (peak_drop + ratchet + regime cooldown) can extract another ~2x compound over a single defensive exit.

Cooldown parameterization is a high-leverage post-loss filter — but small parameter changes can shift fold-by-fold consistency significantly. Always evaluate cooldown sweeps with 3-set seed validation to avoid locking a single-set optimum.

Combining multiple entry styles introduces correlation between them — diversification gains require the styles to fail in different regimes. Quantify cross-attack correlation before combining; otherwise the ensemble just doubles costs.

Confidence weighting tends to underperform hard voting when individual models are mis-calibrated (overconfident in noisy regions) — this matches R070 finding. If you do weighted voting, calibrate the softmax outputs first (temperature scaling) on a validation fold.

Voting parameters interact strongly with cooldown — a stricter vote with shorter cooldown can be equivalent to looser vote + longer cooldown. Always sweep jointly with cooldown to avoid finding a fake local optimum on one axis.

Kelly only works if the probability estimates are calibrated — uncalibrated softmax outputs over-size losing trades because P(TP) is biased high. Calibrate first; then a fractional Kelly (0.25x-0.5x) is the right starting point given parameter uncertainty.

Mean reversion alone fails the worst-fold gate (-42% min alpha), but its uncorrelated payoffs are exactly what V66's selectiveness needs — high trade frequency in regimes where V66 is dormant. A pure mean-reversion strategy is a poor standalone but a great portfolio diversifier (proven in R120).

Tree-based models can't compete with GRU on sequential 15min crypto data — they lack memory across the lookback window. XGBoost trained on summary-stat features could potentially work as a META-model on top of GRU outputs (stacking), but as a primary model it is decisively worse.

Third confirmation (after R062 and R096) that multi-timeframe filters do NOT help once the V66 stack already includes the daily SMA filter. Lesson upgraded to a hard rule: do NOT add more TF filters to V66-class strategies. Future MTF ideas should be tested as the SOLE higher-TF signal, replacing the daily SMA rather than stacking with it.

The diversification thesis works at industrial scale: combining a low-frequency selective momentum strategy (V66) with a high-frequency mean-reversion strategy (MACD_div) in REGIME-AWARE proportions (heavy MACD in trending folds F0/F1/F3, V66-only in mega-bull F2) gives a 5x compound jump WITHOUT sacrificing the worst-fold alpha. Regime-aware beats uniform because each strategy has clear regime preference; the math is purely additive on PnL (no retraining cost). This is the new project record and the most important strategic discovery since R094 (daily SMA filter).

[[classifier-label-leak]] — circular labels invalidate any classifier metric without forward labels.

Inflated by R121 leak. R139 confirmed R128 family ~18% conservative (real winner: R134 τ=0.3).

TA-only strategies don't beat GRU+V66. Need model component.

val_loss != PnL. New features without backtest gain are noise.

[[classifier-label-leak]] [[r128-family-validated]] — forward labels are non-negotiable.

Bar-by-bar simulators are bug magnets. R136 fixed all 4. Foundation for R173 audit.

R135's prior +44% failure was due to 4 stacked bugs: (1) selective trading instead of always-invested, (2) peak_drop on price instead of avg_tp_bar, (3) no trail ratchet (3 levels), (4) wrong V66 params pd_lat=0.38 vs 0.50, pd_bear=0.20 vs 0.42. Closes the reproducibility gap.

Both pillars of R128 family portfolio (V66 + MACD_div) now verified offline. Pure TA strategy, no model retraining. Very fast (1 min CPU). Min α -42 means MACD is a COMPLEMENT (high compound, weak floor), not a replacement for V66.

Diversification effect is real, not just a math artifact. The compound/min α trade-off is fundamental — you can't get both V66's +245 min α AND MACD's compound boost simultaneously with simple fixed weights.

R128 family was approximately correct, not fundamentally broken. The R121b label leak inflated apparent recall but didn't change portfolio direction significantly. No classifier-driven policy with V66 weight < 0.50 keeps min α > 0.

Phase 0 framework = pre-commit gate for any new strategy. Should be wired into CI before next major paradigm experiment.

val_loss of R141 (~0.50) was SAME as canonical. Only the BACKTEST revealed the disaster. Train metrics ≠ deployment performance. Mechanism: V66 thresholds calibrated for canonical-90 distribution; any input filter shift breaks V66 calibration.

Event filter direction CLOSED BOTH WAYS. Canonical 90-percentile is a global optimum. V66 thresholds are TIGHTLY calibrated. Mechanism: tight filter shifts model output distribution to extremes; V66's hardcoded thresholds over/under-trigger.

Labeling change doesn't completely break V66 (kept 4/4 BH) but heavily reduces alpha. tb_vol_10_7 (current canonical) is well-calibrated. Similar lesson as filter: V66 thresholds are tightly coupled to label distribution.

More capacity → overfit. The canonical 2x128 sits at the right point on the bias-variance curve for this data. Confirms V66 is at a tight architecture optimum.

The 'naive' fixed 10% SL is actually optimal given V66's full exit cascade. ATR adapation interacts badly with the other exits (trail, DRSI, peak_drop). Confirms V66 components are co-optimized.

R134 recall is 15-49% (per classifier-label-leak memory). At useful thresholds it filters out too many legitimate V66 trades. Meta-labeling needs a HIGH-RECALL primary classifier, R134 doesn't qualify.

V66 has TAIL RISK in bull (F1) and recovery (F2) regimes — 7-11% probability of negative alpha under permutation. Historic record was favorable but partly luck-dependent. This drove R150 (circuit breaker) and informed deployment risk awareness.

Naive DD circuit breaker is too blunt. V66's existing trail+peak_drop+GRU danger cascade already handles risk. A circuit breaker that triggers on DD that V66 itself caused = double-exit.

BREAKTHROUGH. ADD orthogonal sleeves to V66, DON'T modify V66's inputs/architecture. R151-A captures bear/correction alpha V66 misses without changing V66 itself. No retraining required.

R151-A is path-dependent in F2 (recovery regime). Shorts amplify both upside AND downside. Drives R154 (regime filter) and R160 (friction stress).

Tail risk is inherent to short-during-gap paradigm — moving the threshold redistributes but doesn't eliminate.

Regime filter improves compound but does NOT fix F2 tail risk (the bootstrap fragility is path-dependent, not regime-fragile). F2's recoverable losses cluster in DOWN windows within recovery, not in steady-up periods.

Asymmetry chosen WRONG: k_dn=10 means SL barrier is FARTHER → FEWER SL labels → harder to learn drops. Corrected in R157.

Wrong asymmetry direction. Drives R157 correction.

Improved over R155 in val_loss. Eval in R158.

'Train short-specific GRU' direction CLOSED. Canonical R069 P(SL) remains the BEST short signal. Asymmetric labels make most outcomes Timeout (model can't distinguish TP vs SL).

OOT validation = mandatory gate. Strategies that beat BH in OOT 2026 are not just historical curve-fits.

R151-A is execution-realistic. Half-Kelly sizing is a clean way to reduce variance. Min α drops from +245 to +183 with friction — still robust.

R134 recall too low for sizing application. R148 (filter) and R162 (sizing) both fail for same reason.

Pyramid is the FIRST mechanism to improve min α above V66's +245. F1 (bull) unchanged because gru_safe only fires at event bars (rare in steady bull). The selectivity (3 adds total in 8 years) is FEATURE not bug. Spot-only — no perp needed.

Pyramid CAPTURES BEAR-MARKET INTRADAY RALLIES. 2026 had a sharp rally within the broader bear; pyramid leveraged into it perfectly. Validates the additive paradigm with REAL data.

IN-SAMPLE super-additive because each sleeve fires at different regimes. F0 1322 > max(R163 676, R151-A 880). BUT R167 shows this is path-dependent (see R167).

Super-additivity in backtest is PATH-DEPENDENT. In short OOT windows, sleeves overlap. R163 ALONE is the true winner. Methodology lesson: always validate combined strategies in OOT before deploying.

Multi-timeframe direction CLOSED as standalone retraining.

R163 was sub-optimal. Higher leverage + bigger add + slightly higher trigger improves both compound AND min α. Sweep was worth doing.

Most aggressive variant (higher leverage + bigger add) wins both backtest AND OOT. Pyramid is robust — selectivity (filter on +12% unreal + GRU re-vote) prevents over-trading even with aggressive size.

Instrumentation revealed 4 engine bugs ([[r173-pyramid-was-phantom]]). R170-B was 100% phantom.

Pre-deployment audit prevented deploying phantom strategy. V66 expectations recalibrated. ALL sprint conclusions re-validated.

R151-A is ONLY survivor of sprint. Modest but real edge using canonical R069 P(SL).

V66 thresholds calibrated for canonical GRU softmax. Transformer P(TP) p50=0.18 vs GRU 0.30.

R151-A +2.9pp OOT alpha over V66 is real, survives honest engine + real OOT.

Architecture via threshold tuning CLOSED. Issue is calibration semantics, not just thresholds.

F2 recovery is path-dependent: shorts that win on avg yield losses under permutation. Documented tail risk.

F2 tail risk is path-dependent within F2's own short trades, not from misclassified recoveries. Simple N-day filter useless.

Short WR ~54% throughout (folds vary 50-58%); rolling WR doesn't detect F2-like degradation. Cuts winners.

F2 recovery shorts have lowest WR — F2 PnL comes from longs (62% WR), not shorts (39% WR). Tail risk is structural to F2 shorts.

Audit R173 EMPIRICALLY CONFIRMED. Pyramid direction permanently closed under honest engine. The phantom alpha (155K+ inflated) came 100% from unfunded leverage.

V2 has no inflation because it only uses GRU danger voting (no trail/init_sl/peak_drop). Inflation comes from close-based threshold exits, not from execution price choice per se. V2's pre-audit number was already realistic.

When trail triggers at close[i] (= local min where stop fires), open[i+1] is often ABOVE close[i] (rebound after stop hit) → exec at open[i+1] gives BETTER price than at close[i]. So trail-based strategies tend to DEFLATE under REALISTIC-EXEC, not inflate. V66's 1.66× inflation came from peak_drop, not from trail.

V4 3-regime strategy is honest. The slight min α dip (-15%) is the only delta. Regime-adaptive trails (22/6/10%) work consistently across folds.

V5 V115_cmp = MEJOR producción que V6 V66 incluso honest. ATR-adaptive SL + regime-split trail_tight + peak_drop work synergistically. The +30K min α gap vs V66 comes from V5's tighter SL clip (3-10% vs V66 fixed 10%) catching more loss in bear regimes.

Shorts addon works on V66 (+193 baseline) because adding +1 min α is neutral. On V5 (+266 baseline), shorts pull worst-fold DOWN (-53%). The shorts edge captured by R151-A on V66 is ALREADY captured by V5's tighter cascade (ATR-adaptive SL + peak_drop + regime cooldown). Adding shorts to V5 = double-counting alpha + introducing futures risk for zero net benefit.

Pyramid direction PERMANENTLY closed even under honest accounting. The phantom alpha that R163/R170 showed was 100% from the bug. Initial cap reduction (50%) destroys compound faster than pyramid adds can recover. V5 cascade already captures the upside that pyramid would have harvested.

V5 V115_cmp's existing 5-exit cascade is structurally optimal for its labeling+architecture. Adding more exits causes over-exiting → loses compound without improving min α. Direction CLOSED for marginal tweaks.

V5's '100% always' sizing is structurally optimal. The cascade catches bad trades via exits, not by reducing entry size. Three consecutive negatives (R190 pyramid, R191 exits, R192 sizing) confirm V5 is local optimum for current architecture+labeling+features.

V5's tight init_sl=10% protects against liquidation even at 2× leverage. The cascade exits LONG before drawdown reaches the 25% liquidation threshold. Backtest doesn't model flash-crashes, slippage spikes, API failures — those COULD trigger liquidation in real life. Conservative deployment: 1.25× balances upside (+4× compound) with safety (more buffer to liquidation).

CRITICAL: V5 baseline (+76,725% honest) ASSUMES 0.01% slip per trade. With realistic slip 0.5-2%, compound collapses 99%+ even at 1.0× (no leverage). R195 leverage finding was a MIRAGE that depends on PERFECT execution. Leverage AMPLIFIES the fragility but isn't the root cause — V5's high-frequency cascade is hyper-sensitive to slippage.

V5 baseline already optimal — increasing selectivity removes more GOOD trades than BAD ones. The 'high-freq hurts slip' hypothesis was WRONG. V5 IS robust to plausible slip (0.1% slip still gives +38K/+203). R196 worst-case (2% slip) was unrealistic.

Cascade ensemble doesn't help when one cascade is strictly inferior. V66 was the predecessor; V5 cascade is its improvement. Mixing adds noise. After 8 R&D directions, V5 V115_cmp is uncontested CPU local optimum.

V5's 45% loss rate is IRREDUCIBLE NOISE at entry-time features. The losing trades don't have predictable signatures in (regime, vol, RSI, slope, GRU probs). Meta-labeling direction closed.

Inverse-vol scaling AUTO-HEDGES against slip: high vol periods (worst slip) get smaller positions; low vol (less slip) get bigger. Vol-Kelly maintains multiplicative advantage at all stress levels.