Research architecture freeze · not investment advice

SPY session vol regime

At the close of day t, predict whether session t+1 realizes a low, mid, or high range relative to the trailing 252 sessions known at t.

Method summary

Target: tercile 0 / 1 / 2 of next-session range R = (H-L)/O versus 252-session quantiles known at close t.
Decision time: close of day t. Features use that close and earlier bars only (no same-day H/L/C of t+1).
Validation: expanding walk-forward; train through 2018-12-31; OOS 2019-01-02 to 2026-09-17; retrain every 21 trading days. No shuffled CV.
Bars: multiclass Brier vs climate, beat lagged regime, beat constant mid, beat fitted vol-proxy ablation. Accuracy vs majority is secondary.
Freeze: multinomial logistic 00_logistic_regime_21d. live=false.

Research paper (PDF)

Vol forecast Range freeze Direction null Machine-readable summary (JSON)

Research freeze of spec 00_logistic_regime_21d. Four primary bars pass on the locked walk-forward. Not a live signal. No auto-update. Vol freeze unchanged.

Locked walk-forward · 2019-01-02 to 2026-09-17 · n = 1,938

Headline (logistic vs foils)

Model / foil	Acc	Brier
Logistic freeze	0.5320	0.5677
Climate (OOS class rates)	-	0.6658
Vol-proxy model	0.5212	0.5889
Lagged regime (hard)	0.5243	0.9515
Constant mid	0.3251	1.3498

Class rates: low 0.357 / mid 0.325 / high 0.318. Lagged and constant-mid Brier use hard one-hot rules. The tighter reads are accuracy vs lagged and Brier vs the fitted vol-proxy model.

Primary bars

Bar	Result
Brier vs climate	PASS
Beats lagged regime	PASS
Beats constant mid	PASS
Not just vol	PASS

Multiclass Brier

Lower is better. Climate uses OOS class frequencies. Lagged and mid are hard one-hot foils.

Accuracy by year

Walk-forward logistic, full features. Not a live track.

OOS class mix

Share of low / mid / high labels on the locked OOS sample.