Can we tell at 09:30 whether the RTH range will come in above the median of the last 20 days (relr ≥ 1)?
Criterion met — BSS 17.8 % [12.4 ; 22.9], AUC 74.3 %, accuracy 68.0 % against a base rate of 50.0 %, n = 904 days pooled forward.
Criterion: BSS > 3 and AUC ≥ 60, pooled forward across 2023–2026
| Decision time | 09:30 New York — bars up to t−1 complete, the opening print and every tick of session t before 09:30 |
| Target | relr ≥ 1 — RTH range divided by the median of the 20 previous RTH ranges |
| Data window | 2020-05-29 to 2026-08-14 · 1607 daily bars · last day 2026-08-14 |
| Walk-forward | Dev ≤ 2022 (definitions only), test 2023 / 2024 / 2025 / 2026, expanding with a yearly refit |
| Model | L2 logistic regression on 15 preregistered set C features, standardised on the training window; C from inner validation (the last training year). Features: gap_atr, abs_gap_atr, on_range_atr, ln_on_relr, on_vol_rel, on_delta_norm, open_pos_on, on_ext_recency, pre_range_share, on_high_vs_rh1_atr, on_low_vs_rl1_atr, open_pos_va, prev_ln_relr_rth, prev_ln_relr_mean5, prev_atr5_atr20 |
| Holdout | The PX holdout from 2026-05-01 concerns the PX theses on the tick store. This study is a walk-forward over 2023 to 2026 on daily data — explicitly approved by the user for regime rounds 1 and 2. Every test year is computed from data before it only; no fit ever sees its own test slice. |
Every row is a forecast over the same days. Base rate, majority and persistence are the baselines; the model has to beat them, otherwise it does not count.
| Row | Year | n | Base Rate % | Acc % | AUC % | Brier | BSS % |
|---|---|---|---|---|---|---|---|
| Base Rate | 2023 | 246 | 49.6 | 50.4 | 50.0 | 0.2500 | 0.0 |
| Majority | 2023 | 246 | 49.6 | 50.4 | — | — | — |
| Persistence | 2023 | 246 | 49.6 | 57.7 | 57.5 | 0.2456 | 1.8 |
| Model (LogReg) | 2023 | 246 | 49.6 | 63.4 | 68.8 | 0.2244 | 10.2 |
| Base Rate | 2024 | 250 | 52.8 | 47.2 | 50.0 | 0.2504 | 0.0 |
| Majority | 2024 | 250 | 52.8 | 47.2 | — | — | — |
| Persistence | 2024 | 250 | 52.8 | 59.6 | 59.5 | 0.2409 | 3.8 |
| Model (LogReg) | 2024 | 250 | 52.8 | 69.6 | 75.5 | 0.2019 | 19.3 |
| Base Rate | 2025 | 251 | 47.8 | 47.8 | 50.0 | 0.2501 | 0.0 |
| Majority | 2025 | 251 | 47.8 | 47.8 | — | — | — |
| Persistence | 2025 | 251 | 47.8 | 67.3 | 67.1 | 0.2259 | 9.7 |
| Model (LogReg) | 2025 | 251 | 47.8 | 72.1 | 78.6 | 0.1873 | 25.1 |
| Base Rate | 2026 | 157 | 49.7 | 50.3 | 50.0 | 0.2500 | 0.0 |
| Majority | 2026 | 157 | 49.7 | 50.3 | — | — | — |
| Persistence | 2026 | 157 | 49.7 | 58.6 | 58.3 | 0.2440 | 2.4 |
| Model (LogReg) | 2026 | 157 | 49.7 | 66.2 | 72.0 | 0.2117 | 15.3 |
| Base Rate | pooled | 904 | 50.0 | 48.8 | 49.0 | 0.2501 | 0.0 |
| Majority | pooled | 904 | 50.0 | 48.8 | — | — | — |
| Persistence | pooled | 904 | 50.0 | 61.1 | 59.3 | 0.2386 | 4.6 |
| Model (LogReg) | pooled | 904 | 50.0 | 68.0 | 74.3 | 0.2057 | 17.8 |
Block bootstrap (20-day blocks, 1000 runs, pooled forward): BSS 95 % interval [12.4 ; 22.9].
Confusion matrix pooled: true 0 correct 329 of 452, true 1 correct 286 of 452 (recall 0 = 72.8 %, recall 1 = 63.3 %).
Regularisation chosen per refit — 2023: C = 0.01, n_train = 580 · 2024: C = 3, n_train = 826 · 2025: C = 0.03, n_train = 1076 · 2026: C = 0.1, n_train = 1327.
| Decile | n | mean p % | observed % | difference |
|---|---|---|---|---|
| 1 | 91 | 18.5 | 20.9 | 2.4 |
| 2 | 91 | 28.8 | 25.3 | -3.6 |
| 3 | 91 | 34.5 | 30.8 | -3.7 |
| 4 | 91 | 39.9 | 36.3 | -3.7 |
| 5 | 90 | 44.8 | 45.6 | 0.8 |
| 6 | 90 | 50.3 | 52.2 | 1.9 |
| 7 | 90 | 56.2 | 66.7 | 10.4 |
| 8 | 90 | 62.9 | 58.9 | -4.0 |
| 9 | 90 | 71.7 | 76.7 | 5.0 |
| 10 | 90 | 86.0 | 87.8 | 1.8 |
Sharpness: p ≥ 0.7 on 16.0 % of the days (n = 145), hitting 84.1 % there · p ≤ 0.3 on 16.9 % (n = 153), hitting 79.7 % there.
| Base Rate | the unconditional frequency in the training window, and at the same time the constant comparison forecast. |
| Baseline | the number a model has to beat. Three of them here: base rate, majority (always the more frequent class) and persistence. |
| Persistence | the forecast “today like yesterday” — the value of the same target on the previous day, turned into a rate. |
| BSS | Brier skill score: what percentage of the base rate constant's Brier score the model saves. 0 means equally good, negative means worse. |
| AUC | the probability that a random positive day is scored above a random negative one. 50 is a coin flip. |
| Brier | the mean squared error of the probability. Smaller is better. |
| Sharpness | on what percentage of the days the model says something clear (p ≥ 0.7 or ≤ 0.3) and how often it hits there. A calibrated model without sharpness is useless. |
| Value Area | the price band [rval, rvah] in which 70 % of the RTH volume traded. Always the previous day's here. |
| Trend Day | an RTH with a body ratio ≥ 0.6 — the body of the daily candle fills at least 60 % of its range. |
| Body Ratio | br = |rc − ro| / (rh − rl), computed on the RTH session. |
| relr | RTH range divided by the median of the 20 previous RTH ranges. relr ≥ 1 means: today is bigger than the typical one of the last four weeks. |
| Walk-Forward | every test year is computed from a model that has only seen data before it; the training window grows with every year. No fit sees its own test. |
The numbers in this report are recomputed on every publish — from data/daily-bars-*.json and data/open-features-*.json, which the extractor builds from the tick store.
Run: 1607 daily bars, 0.5 s.