Claude c646e80e35 feat: Vol Forecast & Regime engine (HAR-RV, GJR-GARCH, gradient boosting, HMM) ...

- core/vol_forecast.py: forecasts average variance over the next 1D/1W/1M
  with HAR-RV (weekend dummy for 24/7 crypto), GJR-GARCH(1,1) QMLE with a
  vectorised filter, HistGradientBoosting on HAR/return/calendar features,
  and EWMA as benchmark. Garman-Klass + overnight-gap proxy scaled to
  close-to-close on pre-backtest data only.
- Rolling out-of-sample backtest (refit every 21 days, no look-ahead),
  QLIKE / RMSE / bias per model (Patton 2011), inverse-QLIKE ensemble.
- 2-state Gaussian HMM (Baum-Welch) regime probabilities and durations.
- For BTC/ETH: ensemble forecast vs Deribit ATM implied -> rich/cheap view.
- GET /api/vol/forecast/{pair} (cached 30 min; 503 when history is short
  or sources are down); scikit-learn added (models degrade gracefully
  without it).
- Vol Forecast tab: KPIs, forecast term structure vs implied, HMM regime
  history, OOS backtest vs realised (+DVOL), model leaderboard per horizon.
- 7 new tests on simulations with known answers (proxy bias, GARCH and HMM
  parameter recovery, weekend effect, look-ahead check, endpoint).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FpiqUMAyo4RL5DdQHMWbwE

2026-10-01 05:21:59 +00:00

195 KiB

1440x560px

Raw History