- Add CUDA parallel prefix scan (Blelloch) for EMA: 16 EMAs in 3 batched kernel launches
- Eliminate CPU↔GPU array bouncing; keep arrays on-device until final cache build
- Auto-detect pip-installed CUDA libs for CuPy kernel compilation
- Add --gpu-only and --quick (30-day) CLI flags
- Fix missing _adx_slope in GPUTechnicalEngine (was suppressing transitioning regime)
- Fix ROC early-entry: guard was pos < n+1, should be pos < n+4 to match live engine
- Fix weekend filter: use proper forex weekend (Fri 22:00–Sun 22:00 UTC) instead of weekday >= 5
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
EUR/USD A-grade signals had inverted quality curve (-1.65R) because all
trades got RR≈0.50 and mature_move logic penalized high-score entries.
Fix: add per-pair trade setup params (sl_atr_floor=1.2, min_rr=1.0,
mature_move_threshold=35) and apply quality_scale to both tp_mult and
min_rr so A-grade gets wider TP targets. Also wire EUR config into
backtest via PAIR_PRODUCERS registry.
Result: A-grade now +1.24R (was -1.65R), total EUR/USD +10.06R (was +4.20R).
XAU/USD unchanged at +13.01R.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Track gpu_backtest.py, diagnose_filters.py, and instructions.
Data files (CSVs, parquet cache) remain gitignored.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
harwinder