这个测量准不准、够不够 power(validity)?概率校准、IC 噪声功率地板、模型漂移——对估计量的效度自检
- combined rank-IC = -0.0088(微负,量级远小于月频噪声地板 σ≈0.11)。点估计不构成可检测效应。
- 95% HAC CI = [-0.0336, 0.0159] 跨越零 —— 无法排除「统筹分配」的零效应。
- CI 半宽 0.0336 >> SESOI ±0.01:月频噪声地板下,±0.01 等价结构性不可达(设计级欠功率,非「没效应」)。
- 结论可信非因效应强,而因反泄漏纪律严:PIT 数据 + PurgedGroupKFold+embargo + H6 比特一致。null 是预期的结果。
此叙事严格由已冻结公开数字(IC/CI/p/power-floor)派生,确定性、无外部调用、无副作用。LLM 特征进冻结 OOS = 已知泄漏通道,禁。
效度 · 模型概率可信吗?
校准可靠性 / 自校正显示
leakage-safe · walk-forward 扩窗重拟合 Platt 显示映射
Pooled ECE
0.0348
越低越准(所有 walk-forward 预测汇总)
最新月 ECE
0.1215
最近一个月的样本外校准误差
概率带(最新月)
[0.46, 0.55]
紧绕 base rate ⇒ 模型无判别力(诚实 null)
训练对数(最新月)
30,579
用于重拟合的 realized (score, fwd) 对
Walk-forward calibration reliability (display-only, leakage-safe). For each realized OOS month T, the Platt calibration map is refit on all realized (score, forward-return) pairs STRICTLY before T, then T's tickers are predicted; T's own realized outcome scores the prediction (backward audit). Per-month OOS ECE shows whether the calibrated P(up) stays trustworthy as the realized sample grows; the pooled reliability diagram bins all walk-forward OOS predictions (mean predicted P vs empirical up-frequency). Only the 2-param display map is refit — the frozen LightGBM learner is NEVER touched, and no ledger / frozen surface / research estimator is affected. NOT a research claim and NOT a path to positive rank-IC (research verdict ledger #49: combined rank-IC −0.0088, NULL).
显示层自校正——只重拟合 2 参 Platt 映射,frozen LightGBM 永不触碰;非研究 claim,不制造正 IC(#49 NULL)。