MARKET RESEARCH
命中率與校準
模型停在 2025-08-06 不再更新，拿去考之後的 250 天
目前選取的評估
凍結模型重考，h=5；245 列，期間 2025-09-16 ~ 2026-09-18。
原始方向命中 32.2%；區間涵蓋 73.9% （名目 80%）。方向、機率、區間與交易效用各有不同評估目的，不能互相替代。
本頁整體、方向拆解及校準均受來源與天期篩選控制。 Day 1 政策進度、跨來源校準比較、live 診斷與 live h1 走勢另有固定範圍，標在各卡。
評估來源與分母對帳
source=replay；h=5；期間 2025-09-16 ~ 2026-09-18；245 列／245 個預測日；模型版本 unknown
evaluation_snapshot_id：ef674b253f331fbaee2ccd36f1c0f3c3b0d995124b16b38141ce020df82b9432
| metric_id | 有效 n | 樣本 |
|---|---|---|
| replay.h5.direction_accuracy | 245 | prediction IDsreplay:2025-09-16-h5, replay:2025-09-17-h5, replay:2025-09-18-h5, replay:2025-09-19-h5, replay:2025-09-22-h5, replay:2025-09-23-h5, replay:2025-09-24-h5, replay:2025-09-25-h5, replay:2025-09-26-h5, replay:2025-09-30-h5, replay:2025-10-01-h5, replay:2025-10-02-h5, replay:2025-10-03-h5, replay:2025-10-07-h5, replay:2025-10-08-h5, replay:2025-10-09-h5, replay:2025-10-13-h5, replay:2025-10-14-h5, replay:2025-10-15-h5, replay:2025-10-16-h5, replay:2025-10-17-h5, replay:2025-10-20-h5, replay:2025-10-21-h5, replay:2025-10-22-h5, replay:2025-10-23-h5, replay:2025-10-27-h5, replay:2025-10-28-h5, replay:2025-10-29-h5, replay:2025-10-30-h5, replay:2025-10-31-h5, replay:2025-11-03-h5, replay:2025-11-04-h5, replay:2025-11-05-h5, replay:2025-11-06-h5, replay:2025-11-07-h5, replay:2025-11-10-h5, replay:2025-11-11-h5, replay:2025-11-12-h5, replay:2025-11-13-h5, replay:2025-11-14-h5, replay:2025-11-17-h5, replay:2025-11-18-h5, replay:2025-11-19-h5, replay:2025-11-20-h5, replay:2025-11-21-h5, replay:2025-11-24-h5, replay:2025-11-25-h5, replay:2025-11-26-h5, replay:2025-11-27-h5, replay:2025-11-28-h5, replay:2025-12-01-h5, replay:2025-12-02-h5, replay:2025-12-03-h5, replay:2025-12-04-h5, replay:2025-12-05-h5, replay:2025-12-08-h5, replay:2025-12-09-h5, replay:2025-12-10-h5, replay:2025-12-11-h5, replay:2025-12-12-h5, replay:2025-12-15-h5, replay:2025-12-16-h5, replay:2025-12-17-h5, replay:2025-12-18-h5, replay:2025-12-19-h5, replay:2025-12-22-h5, replay:2025-12-23-h5, replay:2025-12-24-h5, replay:2025-12-26-h5, replay:2025-12-29-h5, replay:2025-12-30-h5, replay:2025-12-31-h5, replay:2026-01-02-h5, replay:2026-01-05-h5, replay:2026-01-06-h5, replay:2026-01-07-h5, replay:2026-01-08-h5, replay:2026-01-09-h5, replay:2026-01-12-h5, replay:2026-01-13-h5, replay:2026-01-14-h5, replay:2026-01-15-h5, replay:2026-01-16-h5, replay:2026-01-19-h5, replay:2026-01-20-h5, replay:2026-01-21-h5, replay:2026-01-22-h5, replay:2026-01-23-h5, replay:2026-01-26-h5, replay:2026-01-27-h5, replay:2026-01-28-h5, replay:2026-01-29-h5, replay:2026-01-30-h5, replay:2026-02-02-h5, replay:2026-02-03-h5, replay:2026-02-04-h5, replay:2026-02-05-h5, replay:2026-02-06-h5, replay:2026-02-09-h5, replay:2026-02-10-h5, replay:2026-02-11-h5, replay:2026-02-23-h5, replay:2026-02-24-h5, replay:2026-02-25-h5, replay:2026-02-26-h5, replay:2026-03-02-h5, replay:2026-03-03-h5, replay:2026-03-04-h5, replay:2026-03-05-h5, replay:2026-03-06-h5, replay:2026-03-09-h5, replay:2026-03-10-h5, replay:2026-03-11-h5, replay:2026-03-12-h5, replay:2026-03-13-h5, replay:2026-03-16-h5, replay:2026-03-17-h5, replay:2026-03-18-h5, replay:2026-03-19-h5, replay:2026-03-20-h5, replay:2026-03-23-h5, replay:2026-03-24-h5, replay:2026-03-25-h5, replay:2026-03-26-h5, replay:2026-03-27-h5, replay:2026-03-30-h5, replay:2026-03-31-h5, replay:2026-04-01-h5, replay:2026-04-02-h5, replay:2026-04-07-h5, replay:2026-04-08-h5, replay:2026-04-09-h5, replay:2026-04-10-h5, replay:2026-04-13-h5, replay:2026-04-14-h5, replay:2026-04-15-h5, replay:2026-04-16-h5, replay:2026-04-17-h5, replay:2026-04-20-h5, replay:2026-04-21-h5, replay:2026-04-22-h5, replay:2026-04-23-h5, replay:2026-04-24-h5, replay:2026-04-27-h5, replay:2026-04-28-h5, replay:2026-04-29-h5, replay:2026-04-30-h5, replay:2026-05-04-h5, replay:2026-05-05-h5, replay:2026-05-06-h5, replay:2026-05-07-h5, replay:2026-05-08-h5, replay:2026-05-11-h5, replay:2026-05-12-h5, replay:2026-05-13-h5, replay:2026-05-14-h5, replay:2026-05-15-h5, replay:2026-05-18-h5, replay:2026-05-19-h5, replay:2026-05-20-h5, replay:2026-05-21-h5, replay:2026-05-22-h5, replay:2026-05-25-h5, replay:2026-05-26-h5, replay:2026-05-27-h5, replay:2026-05-28-h5, replay:2026-05-29-h5, replay:2026-06-01-h5, replay:2026-06-02-h5, replay:2026-06-03-h5, replay:2026-06-04-h5, replay:2026-06-05-h5, replay:2026-06-08-h5, replay:2026-06-09-h5, replay:2026-06-10-h5, replay:2026-06-11-h5, replay:2026-06-12-h5, replay:2026-06-15-h5, replay:2026-06-16-h5, replay:2026-06-17-h5, replay:2026-06-18-h5, replay:2026-06-22-h5, replay:2026-06-23-h5, replay:2026-06-24-h5, replay:2026-06-25-h5, replay:2026-06-26-h5, replay:2026-06-29-h5, replay:2026-06-30-h5, replay:2026-07-01-h5, replay:2026-07-02-h5, replay:2026-07-03-h5, replay:2026-07-06-h5, replay:2026-07-07-h5, replay:2026-07-08-h5, replay:2026-07-09-h5, replay:2026-07-13-h5, replay:2026-07-14-h5, replay:2026-07-15-h5, replay:2026-07-16-h5, replay:2026-07-17-h5, replay:2026-07-20-h5, replay:2026-07-21-h5, replay:2026-07-22-h5, replay:2026-07-23-h5, replay:2026-07-24-h5, replay:2026-07-27-h5, replay:2026-07-28-h5, replay:2026-07-29-h5, replay:2026-07-30-h5, replay:2026-07-31-h5, replay:2026-08-03-h5, replay:2026-08-04-h5, replay:2026-08-05-h5, replay:2026-08-06-h5, replay:2026-08-07-h5, replay:2026-08-10-h5, replay:2026-08-11-h5, replay:2026-08-12-h5, replay:2026-08-13-h5, replay:2026-08-14-h5, replay:2026-08-17-h5, replay:2026-08-18-h5, replay:2026-08-19-h5, replay:2026-08-20-h5, replay:2026-08-21-h5, replay:2026-08-24-h5, replay:2026-08-25-h5, replay:2026-08-26-h5, replay:2026-08-27-h5, replay:2026-08-28-h5, replay:2026-08-31-h5, replay:2026-09-01-h5, replay:2026-09-02-h5, replay:2026-09-03-h5, replay:2026-09-04-h5, replay:2026-09-07-h5, replay:2026-09-08-h5, replay:2026-09-09-h5, replay:2026-09-10-h5, replay:2026-09-11-h5, replay:2026-09-14-h5, replay:2026-09-15-h5, replay:2026-09-16-h5, replay:2026-09-17-h5, replay:2026-09-18-h5 |
| replay.h5.interval_coverage | 245 | prediction IDsreplay:2025-09-16-h5, replay:2025-09-17-h5, replay:2025-09-18-h5, replay:2025-09-19-h5, replay:2025-09-22-h5, replay:2025-09-23-h5, replay:2025-09-24-h5, replay:2025-09-25-h5, replay:2025-09-26-h5, replay:2025-09-30-h5, replay:2025-10-01-h5, replay:2025-10-02-h5, replay:2025-10-03-h5, replay:2025-10-07-h5, replay:2025-10-08-h5, replay:2025-10-09-h5, replay:2025-10-13-h5, replay:2025-10-14-h5, replay:2025-10-15-h5, replay:2025-10-16-h5, replay:2025-10-17-h5, replay:2025-10-20-h5, replay:2025-10-21-h5, replay:2025-10-22-h5, replay:2025-10-23-h5, replay:2025-10-27-h5, replay:2025-10-28-h5, replay:2025-10-29-h5, replay:2025-10-30-h5, replay:2025-10-31-h5, replay:2025-11-03-h5, replay:2025-11-04-h5, replay:2025-11-05-h5, replay:2025-11-06-h5, replay:2025-11-07-h5, replay:2025-11-10-h5, replay:2025-11-11-h5, replay:2025-11-12-h5, replay:2025-11-13-h5, replay:2025-11-14-h5, replay:2025-11-17-h5, replay:2025-11-18-h5, replay:2025-11-19-h5, replay:2025-11-20-h5, replay:2025-11-21-h5, replay:2025-11-24-h5, replay:2025-11-25-h5, replay:2025-11-26-h5, replay:2025-11-27-h5, replay:2025-11-28-h5, replay:2025-12-01-h5, replay:2025-12-02-h5, replay:2025-12-03-h5, replay:2025-12-04-h5, replay:2025-12-05-h5, replay:2025-12-08-h5, replay:2025-12-09-h5, replay:2025-12-10-h5, replay:2025-12-11-h5, replay:2025-12-12-h5, replay:2025-12-15-h5, replay:2025-12-16-h5, replay:2025-12-17-h5, replay:2025-12-18-h5, replay:2025-12-19-h5, replay:2025-12-22-h5, replay:2025-12-23-h5, replay:2025-12-24-h5, replay:2025-12-26-h5, replay:2025-12-29-h5, replay:2025-12-30-h5, replay:2025-12-31-h5, replay:2026-01-02-h5, replay:2026-01-05-h5, replay:2026-01-06-h5, replay:2026-01-07-h5, replay:2026-01-08-h5, replay:2026-01-09-h5, replay:2026-01-12-h5, replay:2026-01-13-h5, replay:2026-01-14-h5, replay:2026-01-15-h5, replay:2026-01-16-h5, replay:2026-01-19-h5, replay:2026-01-20-h5, replay:2026-01-21-h5, replay:2026-01-22-h5, replay:2026-01-23-h5, replay:2026-01-26-h5, replay:2026-01-27-h5, replay:2026-01-28-h5, replay:2026-01-29-h5, replay:2026-01-30-h5, replay:2026-02-02-h5, replay:2026-02-03-h5, replay:2026-02-04-h5, replay:2026-02-05-h5, replay:2026-02-06-h5, replay:2026-02-09-h5, replay:2026-02-10-h5, replay:2026-02-11-h5, replay:2026-02-23-h5, replay:2026-02-24-h5, replay:2026-02-25-h5, replay:2026-02-26-h5, replay:2026-03-02-h5, replay:2026-03-03-h5, replay:2026-03-04-h5, replay:2026-03-05-h5, replay:2026-03-06-h5, replay:2026-03-09-h5, replay:2026-03-10-h5, replay:2026-03-11-h5, replay:2026-03-12-h5, replay:2026-03-13-h5, replay:2026-03-16-h5, replay:2026-03-17-h5, replay:2026-03-18-h5, replay:2026-03-19-h5, replay:2026-03-20-h5, replay:2026-03-23-h5, replay:2026-03-24-h5, replay:2026-03-25-h5, replay:2026-03-26-h5, replay:2026-03-27-h5, replay:2026-03-30-h5, replay:2026-03-31-h5, replay:2026-04-01-h5, replay:2026-04-02-h5, replay:2026-04-07-h5, replay:2026-04-08-h5, replay:2026-04-09-h5, replay:2026-04-10-h5, replay:2026-04-13-h5, replay:2026-04-14-h5, replay:2026-04-15-h5, replay:2026-04-16-h5, replay:2026-04-17-h5, replay:2026-04-20-h5, replay:2026-04-21-h5, replay:2026-04-22-h5, replay:2026-04-23-h5, replay:2026-04-24-h5, replay:2026-04-27-h5, replay:2026-04-28-h5, replay:2026-04-29-h5, replay:2026-04-30-h5, replay:2026-05-04-h5, replay:2026-05-05-h5, replay:2026-05-06-h5, replay:2026-05-07-h5, replay:2026-05-08-h5, replay:2026-05-11-h5, replay:2026-05-12-h5, replay:2026-05-13-h5, replay:2026-05-14-h5, replay:2026-05-15-h5, replay:2026-05-18-h5, replay:2026-05-19-h5, replay:2026-05-20-h5, replay:2026-05-21-h5, replay:2026-05-22-h5, replay:2026-05-25-h5, replay:2026-05-26-h5, replay:2026-05-27-h5, replay:2026-05-28-h5, replay:2026-05-29-h5, replay:2026-06-01-h5, replay:2026-06-02-h5, replay:2026-06-03-h5, replay:2026-06-04-h5, replay:2026-06-05-h5, replay:2026-06-08-h5, replay:2026-06-09-h5, replay:2026-06-10-h5, replay:2026-06-11-h5, replay:2026-06-12-h5, replay:2026-06-15-h5, replay:2026-06-16-h5, replay:2026-06-17-h5, replay:2026-06-18-h5, replay:2026-06-22-h5, replay:2026-06-23-h5, replay:2026-06-24-h5, replay:2026-06-25-h5, replay:2026-06-26-h5, replay:2026-06-29-h5, replay:2026-06-30-h5, replay:2026-07-01-h5, replay:2026-07-02-h5, replay:2026-07-03-h5, replay:2026-07-06-h5, replay:2026-07-07-h5, replay:2026-07-08-h5, replay:2026-07-09-h5, replay:2026-07-13-h5, replay:2026-07-14-h5, replay:2026-07-15-h5, replay:2026-07-16-h5, replay:2026-07-17-h5, replay:2026-07-20-h5, replay:2026-07-21-h5, replay:2026-07-22-h5, replay:2026-07-23-h5, replay:2026-07-24-h5, replay:2026-07-27-h5, replay:2026-07-28-h5, replay:2026-07-29-h5, replay:2026-07-30-h5, replay:2026-07-31-h5, replay:2026-08-03-h5, replay:2026-08-04-h5, replay:2026-08-05-h5, replay:2026-08-06-h5, replay:2026-08-07-h5, replay:2026-08-10-h5, replay:2026-08-11-h5, replay:2026-08-12-h5, replay:2026-08-13-h5, replay:2026-08-14-h5, replay:2026-08-17-h5, replay:2026-08-18-h5, replay:2026-08-19-h5, replay:2026-08-20-h5, replay:2026-08-21-h5, replay:2026-08-24-h5, replay:2026-08-25-h5, replay:2026-08-26-h5, replay:2026-08-27-h5, replay:2026-08-28-h5, replay:2026-08-31-h5, replay:2026-09-01-h5, replay:2026-09-02-h5, replay:2026-09-03-h5, replay:2026-09-04-h5, replay:2026-09-07-h5, replay:2026-09-08-h5, replay:2026-09-09-h5, replay:2026-09-10-h5, replay:2026-09-11-h5, replay:2026-09-14-h5, replay:2026-09-15-h5, replay:2026-09-16-h5, replay:2026-09-17-h5, replay:2026-09-18-h5 |
| replay.h5.calibration | 245 | prediction IDsreplay:2025-09-16-h5, replay:2025-09-17-h5, replay:2025-09-18-h5, replay:2025-09-19-h5, replay:2025-09-22-h5, replay:2025-09-23-h5, replay:2025-09-24-h5, replay:2025-09-25-h5, replay:2025-09-26-h5, replay:2025-09-30-h5, replay:2025-10-01-h5, replay:2025-10-02-h5, replay:2025-10-03-h5, replay:2025-10-07-h5, replay:2025-10-08-h5, replay:2025-10-09-h5, replay:2025-10-13-h5, replay:2025-10-14-h5, replay:2025-10-15-h5, replay:2025-10-16-h5, replay:2025-10-17-h5, replay:2025-10-20-h5, replay:2025-10-21-h5, replay:2025-10-22-h5, replay:2025-10-23-h5, replay:2025-10-27-h5, replay:2025-10-28-h5, replay:2025-10-29-h5, replay:2025-10-30-h5, replay:2025-10-31-h5, replay:2025-11-03-h5, replay:2025-11-04-h5, replay:2025-11-05-h5, replay:2025-11-06-h5, replay:2025-11-07-h5, replay:2025-11-10-h5, replay:2025-11-11-h5, replay:2025-11-12-h5, replay:2025-11-13-h5, replay:2025-11-14-h5, replay:2025-11-17-h5, replay:2025-11-18-h5, replay:2025-11-19-h5, replay:2025-11-20-h5, replay:2025-11-21-h5, replay:2025-11-24-h5, replay:2025-11-25-h5, replay:2025-11-26-h5, replay:2025-11-27-h5, replay:2025-11-28-h5, replay:2025-12-01-h5, replay:2025-12-02-h5, replay:2025-12-03-h5, replay:2025-12-04-h5, replay:2025-12-05-h5, replay:2025-12-08-h5, replay:2025-12-09-h5, replay:2025-12-10-h5, replay:2025-12-11-h5, replay:2025-12-12-h5, replay:2025-12-15-h5, replay:2025-12-16-h5, replay:2025-12-17-h5, replay:2025-12-18-h5, replay:2025-12-19-h5, replay:2025-12-22-h5, replay:2025-12-23-h5, replay:2025-12-24-h5, replay:2025-12-26-h5, replay:2025-12-29-h5, replay:2025-12-30-h5, replay:2025-12-31-h5, replay:2026-01-02-h5, replay:2026-01-05-h5, replay:2026-01-06-h5, replay:2026-01-07-h5, replay:2026-01-08-h5, replay:2026-01-09-h5, replay:2026-01-12-h5, replay:2026-01-13-h5, replay:2026-01-14-h5, replay:2026-01-15-h5, replay:2026-01-16-h5, replay:2026-01-19-h5, replay:2026-01-20-h5, replay:2026-01-21-h5, replay:2026-01-22-h5, replay:2026-01-23-h5, replay:2026-01-26-h5, replay:2026-01-27-h5, replay:2026-01-28-h5, replay:2026-01-29-h5, replay:2026-01-30-h5, replay:2026-02-02-h5, replay:2026-02-03-h5, replay:2026-02-04-h5, replay:2026-02-05-h5, replay:2026-02-06-h5, replay:2026-02-09-h5, replay:2026-02-10-h5, replay:2026-02-11-h5, replay:2026-02-23-h5, replay:2026-02-24-h5, replay:2026-02-25-h5, replay:2026-02-26-h5, replay:2026-03-02-h5, replay:2026-03-03-h5, replay:2026-03-04-h5, replay:2026-03-05-h5, replay:2026-03-06-h5, replay:2026-03-09-h5, replay:2026-03-10-h5, replay:2026-03-11-h5, replay:2026-03-12-h5, replay:2026-03-13-h5, replay:2026-03-16-h5, replay:2026-03-17-h5, replay:2026-03-18-h5, replay:2026-03-19-h5, replay:2026-03-20-h5, replay:2026-03-23-h5, replay:2026-03-24-h5, replay:2026-03-25-h5, replay:2026-03-26-h5, replay:2026-03-27-h5, replay:2026-03-30-h5, replay:2026-03-31-h5, replay:2026-04-01-h5, replay:2026-04-02-h5, replay:2026-04-07-h5, replay:2026-04-08-h5, replay:2026-04-09-h5, replay:2026-04-10-h5, replay:2026-04-13-h5, replay:2026-04-14-h5, replay:2026-04-15-h5, replay:2026-04-16-h5, replay:2026-04-17-h5, replay:2026-04-20-h5, replay:2026-04-21-h5, replay:2026-04-22-h5, replay:2026-04-23-h5, replay:2026-04-24-h5, replay:2026-04-27-h5, replay:2026-04-28-h5, replay:2026-04-29-h5, replay:2026-04-30-h5, replay:2026-05-04-h5, replay:2026-05-05-h5, replay:2026-05-06-h5, replay:2026-05-07-h5, replay:2026-05-08-h5, replay:2026-05-11-h5, replay:2026-05-12-h5, replay:2026-05-13-h5, replay:2026-05-14-h5, replay:2026-05-15-h5, replay:2026-05-18-h5, replay:2026-05-19-h5, replay:2026-05-20-h5, replay:2026-05-21-h5, replay:2026-05-22-h5, replay:2026-05-25-h5, replay:2026-05-26-h5, replay:2026-05-27-h5, replay:2026-05-28-h5, replay:2026-05-29-h5, replay:2026-06-01-h5, replay:2026-06-02-h5, replay:2026-06-03-h5, replay:2026-06-04-h5, replay:2026-06-05-h5, replay:2026-06-08-h5, replay:2026-06-09-h5, replay:2026-06-10-h5, replay:2026-06-11-h5, replay:2026-06-12-h5, replay:2026-06-15-h5, replay:2026-06-16-h5, replay:2026-06-17-h5, replay:2026-06-18-h5, replay:2026-06-22-h5, replay:2026-06-23-h5, replay:2026-06-24-h5, replay:2026-06-25-h5, replay:2026-06-26-h5, replay:2026-06-29-h5, replay:2026-06-30-h5, replay:2026-07-01-h5, replay:2026-07-02-h5, replay:2026-07-03-h5, replay:2026-07-06-h5, replay:2026-07-07-h5, replay:2026-07-08-h5, replay:2026-07-09-h5, replay:2026-07-13-h5, replay:2026-07-14-h5, replay:2026-07-15-h5, replay:2026-07-16-h5, replay:2026-07-17-h5, replay:2026-07-20-h5, replay:2026-07-21-h5, replay:2026-07-22-h5, replay:2026-07-23-h5, replay:2026-07-24-h5, replay:2026-07-27-h5, replay:2026-07-28-h5, replay:2026-07-29-h5, replay:2026-07-30-h5, replay:2026-07-31-h5, replay:2026-08-03-h5, replay:2026-08-04-h5, replay:2026-08-05-h5, replay:2026-08-06-h5, replay:2026-08-07-h5, replay:2026-08-10-h5, replay:2026-08-11-h5, replay:2026-08-12-h5, replay:2026-08-13-h5, replay:2026-08-14-h5, replay:2026-08-17-h5, replay:2026-08-18-h5, replay:2026-08-19-h5, replay:2026-08-20-h5, replay:2026-08-21-h5, replay:2026-08-24-h5, replay:2026-08-25-h5, replay:2026-08-26-h5, replay:2026-08-27-h5, replay:2026-08-28-h5, replay:2026-08-31-h5, replay:2026-09-01-h5, replay:2026-09-02-h5, replay:2026-09-03-h5, replay:2026-09-04-h5, replay:2026-09-07-h5, replay:2026-09-08-h5, replay:2026-09-09-h5, replay:2026-09-10-h5, replay:2026-09-11-h5, replay:2026-09-14-h5, replay:2026-09-15-h5, replay:2026-09-16-h5, replay:2026-09-17-h5, replay:2026-09-18-h5 |
- 無排除
快照識別此來源與篩選後的評估內容；尚非跨檔案原子發布。unknown 代表原帳未保存版本，不推定為目前模型。
事前紀錄 Day 1（2026-08-26 起算）使用者 2026-08-26 決定今天起算，並把既有 8 筆納入
🔴 這份 prereg 是在看過既有 8 筆之後寫的——所以它自己把 written_before_any_data = false 寫在最前面。判準只能來自更早凍結的東西（晉升階梯、五折基線）或純機械推導（ECE 下限），沒有一條可以用「因為那 8 筆長這樣」當理由。
那 8 筆不構成任何證據：模型 8 次全部猜「震盪」，而 n=8 的 ECE 下限是門檻的 2.8 倍。納入它們的意義只有一個——讓計數有一個明確的起點。
| Horizon | 樣本 | 判定 | 依據 | 目前數字 |
|---|---|---|---|---|
| h=1 | 38 | ⏳ insufficient | h=1 只有 38 筆（第一次回看要 60 筆）。只報數字,不下判定——決策規則第 1 條。 | 原始 0.474；平衡 0.479；涵蓋 0.842；ECE 下限 0.1053（門檻的 2.11 倍） |
| h=3 | 35 | ⏳ insufficient | h=3 只有 35 筆（第一次回看要 60 筆）。只報數字,不下判定——決策規則第 1 條。 | 原始 0.457；平衡 0.459；涵蓋 0.829；ECE 下限 0.1260（門檻的 2.52 倍） |
| h=5 | 33 | ⏳ insufficient | h=5 只有 33 筆（第一次回看要 60 筆）。只報數字,不下判定——決策規則第 1 條。 | 原始 0.424；平衡 0.404；涵蓋 0.849；ECE 下限 0.1151（門檻的 2.3 倍） |
判準寫在 config/prereg-2026-08-26-live-day1.toml，由 tx.live_verdict 每天重算——寫在文件裡等人去讀的判準，實際判讀時會變成「看起來還不錯」。 🔴 停止規則：250 筆之後四條判準仍不過，結論是「沒有可交易的 edge」，停止投入，不換特徵重試。
校準樣本與模擬診斷（h=5；跨來源比較，來源篩選不適用）
固定預測機率、以 Bernoulli(p) 模擬結果，得到完全校準假設下的 ECE 分布。 完全校準不等於每次分類都正確；null 中位數也不是誤差下界。
| 資料來源 | 有效配對 n | raw ECE | null 中位數 | null-adjusted diagnostic | 模擬參考 |
|---|---|---|---|---|---|
| 歷史模擬（研究段） | 3,835 | 0.1243 | 0.0128 | 0.1115 | null 中位數低於門檻 |
| 凍結模型重考 ← | 245 | 0.1672 | 0.0515 | 0.1158 | null 中位數為門檻 1.0 倍 |
| 歷史模擬（到最新） | 4,097 | 0.1242 | 0.0128 | 0.1114 | null 中位數低於門檻 |
| 每天真的寫下來 | 33 | 0.0454 | 0.1151 | 0.0000 | null 中位數為門檻 2.3 倍 |
診斷量 = max(0, raw ECE − null 中位數)，零不代表真實校準誤差為零。 固定 10 箱、200 次模擬、seed=0；凍結判準仍使用原始指標，未因本表調整。
🔍 這個訊號真的有用嗎（投資顧問視角）資料：事前 live 紀錄（已結算 h1 38／h3 35／h5 33 筆）——與上方切換的資料來源無關。以下是訊號窗口診斷，尚非可獨立重算的投組績效
退化檢查：預測有分布（最多的「偏多」佔 47%）
它給的範圍，有沒有比「隨手畫一條」還窄？
| 預測天期 | 它給的範圍寬度 | 隨手畫的寬度 | 倍數 | 判定 |
|---|---|---|---|---|
| 1 日 | 3.50% | 4.24% | 0.82 | ✅ 比不做窄 |
| 3 日 | 6.35% | 7.64% | 0.83 | ✅ 比不做窄 |
| 5 日 | 8.26% | 8.87% | 0.93 | ✅ 比不做窄 |
比值大於 1 代表它給的範圍，比「照過去一年的漲跌幅隨手畫一條」還寬——那再漂亮的命中比例，也只是換個好聽的方式說「我不知道」。拿近一年而不是全部歷史當比較基準，是因為它的範圍會跟著近期波動伸縮，用全歷史比會在行情特別平靜或特別亂的時候誤判。
技巧分數（相對「猜多數類」的提升）
| Horizon | n | 命中率 | 基礎率 | skill |
|---|---|---|---|---|
| 1 日 | 38 | 47% | 43% | +0.08 |
| 3 日 | 35 | 46% | 39% | +0.11 |
| 5 日 | 33 | 42% | 38% | +0.07 |
skill ≤ 0 表示不如直接猜多數類。目前每個 horizon 只有 33 筆（第一次判讀要 60 筆），這些數字尚無統計意義——列出來是為了讓它隨閉環累積，而不是現在拿來下結論。
經濟意義（扣一買一賣成本 0.010%，以最新收盤價換算；回測頁寫的 bp 是以 2010–2025 中位價換算，固定費用佔比不同，兩個數字講的是同一套成本參數）
| Horizon | 進場率 | 照訊號做的平均損益 | 一直做多 |
|---|---|---|---|
| 1 日 | 74% | +0.101% | +0.166% |
| 3 日 | 86% | +0.576% | +0.620% |
| 5 日 | 85% | +0.141% | +1.151% |
中性訊號視為空手（不進場、不計成本），所以進場率 0% 時損益恆為 0。「不虧錢」在這裡不等於「有用」——它只代表這段期間模型沒有表態。要判斷這是棄權還是正確迴避，看上方的波動情境：震盪帶被撐寬時，不進場本身就是一個合理的決定。
整體
漲、震、跌分開算的準確率35.3%33.3% 就是丟骰子亂猜的水準（差 +2.0%）
落在預測範圍內73.9%名目水準 80%；須連同寬度、同樣本基準與不確定性評估
筆數2452025-09-16 ~ 2026-09-18 h=5 ⚠️ 紀錄到 2026-09-18，落後資料末端 18 個交易日（record.csv 要手動重跑 build_record.py 才會前進）
不分類別的準確率32.2%只挑最常出現的那一種來猜可得 45.7%
跟事前基準比（replay，h=5，同一批樣本）⚠️ 筆數很少，只能描述、不能判定；基準只用每個預測日當下已成熟的資料（tx.skill），不用測試集自己的比例。歷史段已被看過多次，這張表是診斷，不是未見資料的確認
| 上漲機率（P↑） | Brier（越低越好） | 說明 |
|---|---|---|
| 模型 | 0.2727 | 本頁選取的來源與天期 |
| 事前氣候基準 | 0.2548 | 每個預測日只用當天收盤已成熟的 h 日標籤算上漲比例（平均 37.7%；這批實際 45.7%） |
| 永遠說 50% | 0.2500 | 算術參考，不是這個事件的事前基準 |
Brier skill（相對事前氣候基準）＝ -7.0%（n=245）——模型的上漲機率比「每天都說歷史上漲比例」還差（Brier skill 為負）。
| 80% 區間 | 涵蓋率 | 低於下緣 | 高於上緣 | 平均寬度 | 區間分數（越低越好） |
|---|---|---|---|---|---|
| 模型（q10–q90） | 73.9% | 8.2% | 18.0% | 8.98% | 0.1392 |
| 事前經驗分位（最近 250 筆已成熟） | 66.5% | 15.1% | 18.4% | 7.15% | 0.1428 |
| 條件涵蓋（sigma20 三分位） | n | 模型 | 基準 |
|---|---|---|---|
| 低波動 | 82 | 68.3% | 75.6% |
| 中波動 | 81 | 71.6% | 58.0% |
| 高波動 | 82 | 81.7% | 65.9% |
區間分數 模型較好（0.1392 vs 0.1428，n=245）。區間分數把寬度與漏抓一起計價：開很寬也涵蓋得到 80%，但分數會變差，所以「涵蓋率接近 80%」本身不證明能力。
這個 horizon 的彙總（h=5；要比較 1／3／5 日請用上方切換）
| Horizon | 筆數 | 方向命中率 | 區間涵蓋率 | 平均報酬 |
|---|---|---|---|---|
| 5 日 | 245 | 32.2% | 73.9% | +1.238% |
拆開看：說偏多／中性／偏空時各自準不準
| 預測方向 | 筆數 | 方向命中率 | 平均報酬 |
|---|---|---|---|
| 偏空 | 92 | 21.7% | +1.387% |
| 中性 | 114 | 36.0% | +0.916% |
| 偏多 | 39 | 46.2% | +1.824% |
隔天預測的猜對率走勢（真的寫下來的紀錄：已對答案 38 筆／38 天）
圓點 = 那天對或錯（1 或 0），虛線 = 到目前為止的平均， 實線 = 滾動平均。 只畫「隔天」這一種——五天後的難度不同，混在一起看會被「當天剛好 輪到哪一種到期」帶著跳。目前筆數還很少，任何看起來的趨勢都只是巧合： 這張圖現在的用途是確認系統每天有在累積，不是判斷模型好不好。 累積到 60 次時才用事先寫好的標準判定。
說幾成漲，就真的有幾成漲嗎
淺色 = 預測上漲機率，深色 = 實際上漲比例。 raw ECE 0.1672（既有門檻 < 0.05）｜Brier 0.2727。
[0.0, 0.1)（n=11）
預 8%／實 45%
[0.1, 0.2)（n=34）
預 16%／實 32%
[0.2, 0.3)（n=92）
預 27%／實 38%
[0.3, 0.4)（n=77）
預 34%／實 58%
[0.4, 0.5)（n=20）
預 44%／實 50%
[0.5, 0.6)（n=7）
預 52%／實 43%
[0.6, 0.7)（n=4）
預 65%／實 75%
有效校準配對 245 筆；10 個等寬箱，200 次 Bernoulli(p) 模擬，seed=0。null 中位數 0.0515、95 分位 0.0839。null-adjusted diagnostic = 0.1158，它是模擬參考量，不能推論真實誤差為零或模型通過。
Brier 同時受校準、判別力與事件基礎率影響。固定 0.5 預測的 0.25 只是一個算術參考，
不是這個上漲事件的事前基礎率基準。校準分數不能單獨證明預測或交易能力。