# ETH pivot picker v4 — Round B sealed evaluation Closed-bar localization only. No fees, PnL, leverage or trade rules. Both operating points below were chosen on burned development before the single opening. Every TF is reported even when weak. Search: **1152 new configs**, 3360 trials in the familywise correction; shifted-null p=0.050 (19 shifts). This tests localization against rotated calls, not superiority to baselines. Freeze SHA-256 `0ae7017e27cfbb4ff4b0aa4219611bbdf6ed5f787ebc5a04fb060c06ecd60dae`. Single opening: 2026-11-12T19:56:46.159435+00:00. Full numeric results, curves, baselines, reliability bins and development selection records are in `results_v4.json`; flat live parameters are in `settings_v4.json`. ## 1 minute Best frozen candidate at its trained scale on the final (descriptive): **event_ensemble `838d92ef5b08b502`**, ['atr', 2]; W3 excess 43.8%. This does not change the frozen live settings. | Frozen point / target | Precision W1 / W3 / W5 | Recall W3 | Median / p90 error bp | Median / p90 error ATR | Lag median / p10 / p90 bars | Calls / false calls per day | |---|---|---|---|---|---|---| | high_precision / ['atr', 3] | 51.6% / 59.7% / 67.0% | 1.5% | 17.1 / 39.9 | 1.709 / 2.783 | 0.0 / -2.0 / 0.0 | 2.19 / 0.88 | | balanced / ['atr', 2] | 43.5% / 52.8% / 57.2% | 22.3% | 16.4 / 43.0 | 1.691 / 2.894 | 0.0 / -2.0 / 2.0 | 73.60 / 34.70 | | Point | Brier / base-rate Brier | Log loss | PR-AUC / base-rate lift | Precision lift vs best baseline / shifted null | |---|---|---|---|---| | high_precision | 0.2048 / 0.1630 | 2.8268 | 0.2059 / 1.00 | 1.18 / 2.86 | | balanced | 0.3946 / 0.2418 | 5.4028 | 0.4275 / 1.04 | 0.74 / 1.77 | **Assessment.** Full-bar Brier is worse than the constant base-rate forecast at both frozen points. The sparse callers localize some extremes but leave most pivots uncalled; a zero probability on those bars is overconfident. The point names describe development selection objectives, not guarantees: high_precision maximizes excess over the shifted null, balanced maximizes excess × √calls. Targets may differ, so raw precision alone does not rank the two points. Nested development: 18 outer rows; **15 eligible, 3 excluded** from mean excess 32.8%. Exclusions reflect the 30-call floor, which can penalize sparse baselines on 15m. Outer labels are purged at confirmation time; each block is trimmed by one day plus five bars. All six target scales and thresholds 0/.5/.65/.8/.9 are scored, with W1/3/5 one-to-one precision@recall curves. Threshold zero means all emitted candidates, not every bar. Reliability curves include no-call zeros. Zones are heuristic ±0.5 ATR around the call high/low. The 500-minute post-v3 segment is recorded separately; insufficient warm-up is not a zero result.
## 3 minute Best frozen candidate at its trained scale on the final (descriptive): **event_ensemble `6931682922d98743`**, ['atr', 3]; W3 excess 47.9%. This does not change the frozen live settings. | Frozen point / target | Precision W1 / W3 / W5 | Recall W3 | Median / p90 error bp | Median / p90 error ATR | Lag median / p10 / p90 bars | Calls / false calls per day | |---|---|---|---|---|---|---| | high_precision / ['atr', 3] | 50.2% / 59.4% / 65.9% | 6.0% | 39.1 / 99.0 | 1.755 / 3.376 | 0.0 / -2.0 / 1.0 | 2.66 / 1.08 | | balanced / ['fixed', 0.003] | 46.8% / 63.3% / 68.3% | 43.7% | 18.9 / 60.9 | 0.682 / 2.136 | -1.0 / -3.0 / 1.0 | 65.44 / 24.03 | | Point | Brier / base-rate Brier | Log loss | PR-AUC / base-rate lift | Precision lift vs best baseline / shifted null | |---|---|---|---|---| | high_precision | 0.1902 / 0.1547 | 2.6205 | 0.1949 / 1.02 | 1.06 / 3.13 | | balanced | 0.5143 / 0.2469 | 6.9728 | 0.5825 / 1.05 | 0.74 / 1.49 | **Assessment.** Full-bar Brier is worse than the constant base-rate forecast at both frozen points. The sparse callers localize some extremes but leave most pivots uncalled; a zero probability on those bars is overconfident. The point names describe development selection objectives, not guarantees: high_precision maximizes excess over the shifted null, balanced maximizes excess × √calls. Targets may differ, so raw precision alone does not rank the two points. Nested development: 18 outer rows; **18 eligible, 0 excluded** from mean excess 38.5%. Exclusions reflect the 30-call floor, which can penalize sparse baselines on 15m. Outer labels are purged at confirmation time; each block is trimmed by one day plus five bars. All six target scales and thresholds 0/.5/.65/.8/.9 are scored, with W1/3/5 one-to-one precision@recall curves. Threshold zero means all emitted candidates, not every bar. Reliability curves include no-call zeros. Zones are heuristic ±0.5 ATR around the call high/low. The 500-minute post-v3 segment is recorded separately; insufficient warm-up is not a zero result.
## 5 minute Best frozen candidate at its trained scale on the final (descriptive): **event_ensemble `385c20ca30c7eb28`**, ['fixed', 0.006]; W3 excess 43.7%. This does not change the frozen live settings. | Frozen point / target | Precision W1 / W3 / W5 | Recall W3 | Median / p90 error bp | Median / p90 error ATR | Lag median / p10 / p90 bars | Calls / false calls per day | |---|---|---|---|---|---|---| | high_precision / ['fixed', 0.006] | 62.7% / 75.0% / 79.2% | 6.7% | 46.1 / 118.9 | 1.411 / 3.207 | 0.0 / -2.0 / 0.0 | 2.79 / 0.70 | | balanced / ['fixed', 0.006] | 37.5% / 49.1% / 52.8% | 38.2% | 34.1 / 99.5 | 0.876 / 2.485 | -1.0 / -3.0 / 1.0 | 24.52 / 12.48 | | Point | Brier / base-rate Brier | Log loss | PR-AUC / base-rate lift | Precision lift vs best baseline / shifted null | |---|---|---|---|---| | high_precision | 0.3355 / 0.2239 | 4.6266 | 0.3435 / 1.01 | 1.02 / 2.39 | | balanced | 0.3207 / 0.2239 | 4.3208 | 0.3695 / 1.09 | 0.67 / 1.91 | **Assessment.** Full-bar Brier is worse than the constant base-rate forecast at both frozen points. The sparse callers localize some extremes but leave most pivots uncalled; a zero probability on those bars is overconfident. The point names describe development selection objectives, not guarantees: high_precision maximizes excess over the shifted null, balanced maximizes excess × √calls. Targets may differ, so raw precision alone does not rank the two points. Nested development: 18 outer rows; **17 eligible, 1 excluded** from mean excess 34.9%. Exclusions reflect the 30-call floor, which can penalize sparse baselines on 15m. Outer labels are purged at confirmation time; each block is trimmed by one day plus five bars. All six target scales and thresholds 0/.5/.65/.8/.9 are scored, with W1/3/5 one-to-one precision@recall curves. Threshold zero means all emitted candidates, not every bar. Reliability curves include no-call zeros. Zones are heuristic ±0.5 ATR around the call high/low. The 500-minute post-v3 segment is recorded separately; insufficient warm-up is not a zero result.
## 15 minute Best frozen candidate at its trained scale on the final (descriptive): **event_ensemble `c613559b384f41ad`**, ['atr', 2]; W3 excess 30.6%. This does not change the frozen live settings. | Frozen point / target | Precision W1 / W3 / W5 | Recall W3 | Median / p90 error bp | Median / p90 error ATR | Lag median / p10 / p90 bars | Calls / false calls per day | |---|---|---|---|---|---|---| | high_precision / ['atr', 2] | 52.1% / 61.5% / 66.6% | 12.9% | 66.5 / 175.3 | 1.331 / 2.745 | 0.0 / -2.0 / 1.0 | 2.19 / 0.84 | | balanced / ['fixed', 0.006] | 52.1% / 68.6% / 73.6% | 40.3% | 44.3 / 129.3 | 0.714 / 2.088 | -1.0 / -2.0 / 1.0 | 13.77 / 4.33 | | Point | Brier / base-rate Brier | Log loss | PR-AUC / base-rate lift | Precision lift vs best baseline / shifted null | |---|---|---|---|---| | high_precision | 0.3625 / 0.2328 | 4.9842 | 0.3780 / 1.02 | 0.89 / 1.99 | | balanced | 0.6008 / 0.2268 | 8.1988 | 0.6745 / 1.03 | 0.79 / 1.34 | **Assessment.** Full-bar Brier is worse than the constant base-rate forecast at both frozen points. The sparse callers localize some extremes but leave most pivots uncalled; a zero probability on those bars is overconfident. The point names describe development selection objectives, not guarantees: high_precision maximizes excess over the shifted null, balanced maximizes excess × √calls. Targets may differ, so raw precision alone does not rank the two points. Nested development: 18 outer rows; **9 eligible, 9 excluded** from mean excess 24.4%. Exclusions reflect the 30-call floor, which can penalize sparse baselines on 15m. Outer labels are purged at confirmation time; each block is trimmed by one day plus five bars. All six target scales and thresholds 0/.5/.65/.8/.9 are scored, with W1/3/5 one-to-one precision@recall curves. Threshold zero means all emitted candidates, not every bar. Reliability curves include no-call zeros. Zones are heuristic ±0.5 ATR around the call high/low. The 500-minute post-v3 segment is recorded separately; insufficient warm-up is not a zero result.
## Limits and review disposition - All v1-v3 holdouts and v3 2024 final are burned development only. The sealed 2025 year is backwards-time and Binance-to-OKX venue transfer, not forward validation. - Round A archive fields were compacted after its search; inherited manifest discloses that transport operation. No final scoring or selection used prices before this opening. - Final coverage: 525600 ETH and 525600 BTC-close 1m bars in 2025; post-v3 ETH is only the inherited sealed 500 minutes, not bars arriving after the seal timestamp. - BTC final history contains close only and is used as context; BTC pivot performance is not scored. - No historical liquidation, order book, options, open interest or long-short feed was acquired. No scored claim depends on live-only data. - 19 common circular shifts give p resolution 0.05; null stationarity is an approximation. Familywise significance is not proof of improvement over baselines. - Finalists and both operating points were frozen before opening; best-on-final ranking is descriptive and does not replace deployed development-selected settings. - No price-zone coverage guarantee. Small samples and missing rows remain explicit; probabilities target per-bar proximity, which differs from one-to-one pivot matching. All three supplied findings reproduced and were fixed with regression tests: modal finalist ties now use eligible outer excess, reversal state allows only the first reversal per run, and nested excess means disclose included/excluded rows. The named review_v4a.md was not supplied; the three inline findings are the available review. Historical Round A metrics are retained as historical trials, while Round B re-evaluates corrected calls. Offline reproduction and parity checks: see `lab/v4-round-b.md`.