mindesbunister a669058636 docs: V11 progressive sweep results - 1,024 configs complete ...

SWEEP COMPLETED: 33.2 minutes, 4 workers, ALL 1,024 configs tested

KEY FINDINGS:
 NO zero-signal configs (flip_threshold fix successful)
 Top strategy: 1.97 PF, 74.7% WR, $2,416 PnL (766 trades)
 5× better P&L than v9 baseline ($405 → $2,416)
 96% less drawdown than v9 (-$1,360 → -$55)

CRITICAL ANOMALY DISCOVERED:
 flip_threshold=0.35/0.40 generating 3-4× FEWER signals than expected
  - flip=0.30: 1,271 avg signals (Worker1) ✓
  - flip=0.35: 304 avg signals (Worker2) ⚠️
  - flip=0.40: 276 avg signals (Worker2) ⚠️
  - flip=0.45: 920 avg signals (Worker1) ✓

Expected: 0.30 > 0.35 > 0.40 > 0.45 (linear decrease)
Actual: 0.30 (1,271) > 0.45 (920) > 0.35 (304) > 0.40 (276)

Possible causes:
1. Indicator bug in mid-range flip detection
2. Worker2 deployment issue (stale code?)
3. Dataset artifact (2024 SOL specific pattern)

OPTIMAL PRODUCTION CONFIG:
- flip_threshold=0.45 (all top 10 use this)
- adx_min=15 (strictest filter, all top 10)
- long_pos_max=95, short_pos_min=5 (permissive)
- vol_min=0.0 (no volume filter)
- RSI parameters DON'T MATTER (identical results)

ADX FILTER VALIDATION:
 adx=0: 1,162 signals (most, as expected)
 adx=5: 582 signals (50% reduction)
 adx=10: 572 signals (similar to adx=5)
 adx=15: 455 signals (least, as expected)

NEXT STEPS:
1. Investigate flip=0.35/0.40 anomaly (re-run on Worker1)
2. Forward test flip=0.45, adx=15 config on 2025 data
3. Deploy to production if validation passes

Files:
- cluster/V11_SWEEP_RESULTS.md (comprehensive analysis)
- cluster/v11_results/*.csv (local copies of all 4 chunks)

2025-12-07 00:34:49 +01:00

1.9 KiB

Raw Blame History

View Raw