b793a8114b
fix: adaptive exploration boost when SOTA dominated by single indicator
...
- +25% explore when >80% SOTA shares same indicator
- Force non-dominant indicator every 100 iterations
- Base exploration raised to 40% (effective 30% with 20 SOTA)
2026-05-31 17:43:45 +02:00
6bce4f2405
feat: multi_role strategy — trend filter + entry gating across TFs
...
- New strategy type: trend_ind(higher TF) → entry_ind(lower TF)
- Entry only fires when trend confirms direction (directional gating)
- 15² × 3×2 = 1,350 indicator/TF combinations
- Found: MACD(30min)→ADX(15min) = Sharpe 102.37, +32.2%/month
2026-05-31 17:39:48 +02:00
2e028ffc1e
fix: raise exploration rate to 30% — discover indicators beyond MACD
2026-05-31 17:28:44 +02:00
6cd362aa25
feat: R&D loop — Optuna optimization + LightGBM ML training
...
- Optuna: every 500 iterations, 20-trial hyperparameter optimization
- ML: every 2000 iterations, LightGBM classifier on SOTA indicator signals
- Numba backtest: 245× faster (735M bars/s)
- All 3 discovery methods: explore → exploit → optuna → ml
2026-05-31 17:21:23 +02:00
a373710454
perf: Numba GPU-accelerated backtest — 245× faster (735M bars/s)
...
- Replaced vbt_backtest with Numba JIT-compiled bar-by-bar simulation
- 2.26M bars in 0.003s (was 0.74s)
- 50,000 iterations now 2.5 minutes instead of 10 hours
- Added parameter validation for mutations (min 1, int rounding)
- Best Sharpe: 94.89 (ROC) — 28% monthly
2026-05-31 17:06:07 +02:00
ee3d7786c3
feat: new R&D loop — indicator discovery with exploit/explore mechanics
...
- Replaces broken factor pipeline with working indicator-based loop
- Architecture: hypothesize → evaluate → feedback → record
- Bandit-inspired: 70% exploit (mutate best), 30% explore (random)
- 15 indicators, 3 strategy types (single, multi-tf, portfolio)
- 300 iterations in 215s — discovered MACD 4-TF at Sharpe +28.93
- Adaptive exploration rate (30%→10% as SOTA grows)
- Autonomous improvement: SAR(+16)→SAR(+22)→MACD(+25)→MACD(+28)
2026-05-30 12:48:39 +02:00
TPTBusiness