branch main commit 4fdecb8 built last · feat(ideas): replace git-based ideas with Cloudflare KV backend
How to read this table: Each row documents why a hypothesis didn't work and under what FUTURE condition it would be worth retrying. If you have a new idea, grep this page first — if it's already here, save the Colab hours.
RoundDateParadigmHypothesis Predicted ΔActual ΔFailure mode Re-explore triggerSalvage
R119 mtf-filter 4h EMA/RSI/slope as a no-trade filter on V66 avoids whipsaws +5-10% compound, MDD ↓ All 5 filter variants HURT; MTF_all strictly worse (-119 min α) noise Only with a LEARNED MTF gate (RL or attention), not hand-coded Confirms R096 — MTF context belongs in features, not in gates
R118 alt-model XGBoost gives independent alpha vs NN, ensemble diversity +10-20% min α via decorrelation min α -128% (0/4 beats B&H); compound +245% (flat) regime-fragility Only with sequence-aware features (lagged returns matrix) Feature importances confirm clean 36-feature set is dominant
R116 sizing Fractional Kelly / vol-scaled / DD-throttled reduces MDD MDD ↓ 5-10pp at modest compound cost ALL 9 configs lose compound; 100% per trade optimal regime-fragility + engine bug (MDD all 100%) ONLY after fixing equity tracking bug Engine bug documented; vol_scaled revisitable post-fix
R115 other Asymmetric voting threshold beats default (3,3) +5-10% compound via tighter danger detection (3,3) confirmed global max across 25 configs marginal Only if voting mechanism changes (confidence-weighted) Voting threshold is solved — no further sweeps needed
R114 ensemble Confidence-weighted voting > binary ≥3/5 +10-15% compound via better signal extraction ALL 21 configs lose -99% compound marginal (binary already extracts salient signal) Only if ensemble composition changes (GRU+LSTM+Tformer) Confirms voting robust; future ensemble work skips this dim
R113 hybrid-sma Multi-attack (A1 cooldown + A2 dual-regime + A5 exit-AND) +25-35% compound, +245% min α maintained ALL 8 configs REJECTED; baseline V66 wins all regime-fragility (Agent B over-predicted) Only with new regime detection (not ATR-based) Maps V66 hyperparameter landscape — sharp local max confirmed
R112 exit-logic Sweep cd_loss_uniform around 48 to find better cooldown Marginal improvement via fine tuning Baseline (48) confirmed optimal marginal — V66 at sweet spot Only with new exit logic (not just duration) V66 hyperparameters exhausted → paradigm shift needed