fomc-rv-qlike-forecaster

Deployable SPX intraday realized-vol forecaster (SWE 599 thesis): 5-seed QLIKE-trained deep ensemble, HAR-residual-stacked, beats HAR on QLIKE at h=1/5/22 (block-confirmed), with walk-forward OOS conformal bands.

Files

• production_artifact.json: per-horizon har_coef [int,daily,weekly,monthly], feat_mean/feat_std, resid_mean/resid_std, conformal_quantiles (alpha->log-RV band half-width), seed_state_dicts, feature_order.
• h{1,5,22}_seed{11,22,33,44,55}.pt: per-horizon/seed QLIKE-DLq residual-head state_dicts.
• production_eval.json: walk-forward QLIKE/R2 + band coverage.

Forecast (per horizon h)

1. features in feature_order from latest realized measures (HAR lags + RS+/RS-/BV/RQ/skew/kurt/Parkinson/log-volume).
2. HAR floor = har_coef . [1,daily,weekly,monthly].
3. standardize feats; run each seed head; average standardized residuals; de-standardize.
4. point log-RV = HAR + ensemble residual; variance = exp(log-RV).
5. band(1-alpha) = [exp(point-q), exp(point+q)], q=conformal_quantiles[alpha].

Serving needs the _build_model class from the fed-pulse backend. Inputs: dataset fomc-text-volatility-data.