R&D Pipeline
Queued: prioritized by impact / (cost × risk). In-flight: max 3 at once. Done: last 5 outcomes for context.
Queued 8
Paper trade R128d 'soft+25% floor' as V7 sleeve on Binance testnet
R128d compound 130,333% (+84% V66) min α 30.4 4/4 BH is current best diversification candidate. Validate live execution 2 weeks before promoting V7 alongside V66.
Better V66-zone labels — use actual V66/MACD daily PnL
R121b's binary label is heuristically derived from features. Replacing with TRUE daily V66 vs MACD outperformance label could further improve recall and close gap toward min α 245.
Relaxed event filter sweep (notebook ready, GPU)
Tag r122 ready for Colab. Tradeoff curve: how loose can the event filter go before min α collapses? 10 configs sweep, ~17 GPU-h.
Multi-horizon GRU ensemble (5min + 15min + 1h)
Each horizon sees different patterns; union of trade calendars → 2.5-3× V66 trades. Heavy GPU (12+ hours A100).
Phase 0 validation infrastructure (src/validation/)
Without it, no candidate can be safely promoted — blocks the entire promotion pipeline
RL agent (PPO/DQN) — Stable-baselines3
Unexplored paradigm — no R&D rounds yet, completely orthogonal to GRU
Sentiment / LLM features from crypto news
Idea 002 — non-technical alpha source, complementary to GRU's price-only features
Transformer / TFT architecture
Last-tier exploration — only if all simpler paths plateau
In-flight 0
No experiments in flight. R120 ready to start.
Recently done 12
Portfolio with R121b binary classifier (best of R128 series)
BEST candidate yet — R121b 'soft + 25% floor': compound 130,333% (+84% V66), min α 30.4, 4/4 BH. Binary classifier with 100% F2 V66-zone recall closes part of gap.
Binary classifier — is_v66_zone (GBM, 14 features)
PROMOTE — F2 V66-zone recall = 1.000 (vs R121's 0.852, +14.8pp), F1 recall 0.961, avg WF accuracy 94.5%. Unlocks R128d portfolio improvement.
TA + daily SMA filter (last TA attempt)
Even with R094's most-impactful component (daily SMA filter), 6 TA variants all fail. Best macd_div_daily: compound 305%, min α -189. TA paradigm fully closed.
TA with proper divergence + tight exits
9 variants tested with proper divergence detection + MACD-cross exits. ALL negative compound. Implementation matters but result robust: TA-only doesn't work.
10 multi-TA combination strategies (standalone)
All 10 standalone TA combos (MACD+vol, BB squeeze, RSI oversold+vol, Donchian breakout, OBV div, VWAP, triple confluence) failed. None beats V66.
Proba-weighted allocation sweep (11 policies)
BEST: Policy G (P_rec>0.2 OR P_bear>0.4 OR P_lat>0.5 → 100% V66, else 25/75) → +30% compound (91,856%) vs V66, min α 138.6 (vs 21.5 in R128b hard). 4/4 BH. Pareto improvement.
Portfolio deployment backtest (R121 per-bar switch)
DIVERSIFY — 25/75 default → +35% compound vs V66, 4/4 BH, but min α 21 (V66=245). Pareto sweep showed classifier 85% recall is too low for strict replacement. Deploy as V7 sleeve, not replacement.
Real-time regime detector (4-class RF)
PROMOTE — 90.8% overall acc, 85.2% min critical-class (bear/bull/recovery >=80% all folds), 0.042ms/bar latency. Unlocks R128 portfolio deploy.
Portfolio V66 + MACD_div (regime-aware)
7 configs PASS — top: 362,991% compound (5.1× V66), min α matches 245, 16× more trades. Required regime detector (→ R121 done).
Multi-timeframe filter (4h context)
All 5 filters degrade V66; best (no_extreme_RSI) +168 α vs +245 baseline
XGBoost alternative model
+245% compound vs +70,576% V66 — trees crushed by GRU 287×; 0/4 beats B&H
Mean reversion strategies portfolio
MACD_div +179K% compound BUT min α -42 — became R120 portfolio candidate