Research architecture freeze · not investment advice

SPY session range expansion

At the close of day t, predict whether session t+1’s range R = (H-L)/O exceeds the mean of the prior ten session ranges.

Method summary

Target: whether next-session range R = (H-L)/O exceeds the mean of the prior ten session ranges.
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 2026-07-31; retrain every 21 trading days. No shuffled CV.
Bars: Brier vs climate, beat persistence, beat RV-only, Spearman of P against realized R. Accuracy vs majority is secondary.
Freeze: LightGBM trial 22 (01_optuna_trial22_lgbm_21d). Named baseline 00_lgbm_params_2026_08_01 is retained. live=false.
Paper hold-out: each new close after 2026-07-31 is scored under the frozen recipe into a ledger only. Independent of the vol workflow. Not a live badge. Public panel below is n and Brier only.
Live later, if at all: evaluate that hold-out first. Then shadow on new closes and dual-control. Range must not be able to take vol down.

Research paper (PDF)

Vol forecast Regime freeze Direction null

Research freeze of spec 01_optuna_trial22_lgbm_21d. Not a live signal. No auto-update. Vol freeze unchanged. A promotion gate exists and cannot write live.

Locked walk-forward · 2019-01-02 to 2026-07-31 · n = 1,905

Headline (frozen tree vs named baseline)

Metric	2026-08-01 baseline	Frozen trial 22
Brier	0.21675	0.21378
Accuracy	0.6451	0.6514
Spearman(P, R)	0.1542	0.1642
Climate Brier	0.2454	0.2454

Four primary bars all pass: Brier vs climate, persistence, RV-only ablation, Spearman. RV-only on the frozen hyper-parameters is tied on accuracy (0.6514) and worse on Brier (0.2191).

Paper hold-out · after 2026-07-31 · not live

Paper hold-out (frozen recipe, new closes)

Origins after the freeze cut are scored into a ledger only. This is not a live badge and is not mixed into the 1,905-row table above.

Metric	Paper hold-out
Scored origins	32 of 33
Origin window	2026-07-31 to 2026-09-16
Brier	0.2448
Climate Brier	0.2148
Spearman(P, R)	-0.153

On n=32 the frozen recipe does not beat climate Brier, and Spearman is negative. Sample is still small. Keep the research freeze. Do not promote.

Promotion gate (last recorded 2026-08-20)

PSI: last 63 days vs prior 252, on freeze scores and stationary features (range_z_20, atr10_ratio). Raw range/RV levels are reported only.
Golden master: 1,000 static OOS rows; freeze self-check correlation 1.0.
Shadow: 48 hours. Recorded decision WAIT (clock started on that run).
Dual control: two distinct actors.
Live write: always refused.

Recorded PSI tier: CAUTION. GraphForge encodes those fail-closed transitions in the range repo (rangeforge/).