效度 · 模型在漂移吗?

模型健康 / 漂移监测

leakage-safe · 最近 6 个实现月 vs 历史

us中度漂移
PSI 评分分布漂移(最近 vs 历史) · PSI = 0.226 · nrecent 2934 / nhist 28134

PSI

0.226

IC 最近

-0.002

IC 全程

0.004

base rate

0.54→0.52

cn显著漂移
PSI 评分分布漂移(最近 vs 历史) · PSI = 0.542 · nrecent 1795 / nhist 17960

PSI

0.542

IC 最近

0.081

IC 全程

-0.046

base rate

0.45→0.40

模型卡
冻结确认模型的可追溯档案:身份 · 数据 · 模型 · 评估协议 · 结果 · 治理——全部机器读自冻结工件并哈希钉定。

身份

字段值
阶段(ledger 行)B
config_sig17245a75d2d4cd17c68f36a9d0f4b4f7f3baf1db31b33b87be79e6e4a11400da
结果目录runs/results/17245a75d2d4cd17c68f36a9d0f4b4f7f3baf1db31b33b87be79e6e4a11400da
Aionis 版本0.1.0
schema1
H6 确定性PASS
run 时间(工件 ts,非时钟)2026-07-28T04:27:39+00:00

预期用途

字段值
主张Display-layer documentation of the frozen confirmatory model: a pre-registered, two-tailed, null-expected rank-IC claim whose config was committed to the append-only ledger before any out-of-sample metric was observed. The card exists for transparency and reproducibility inspection, not for performance advertisement.
仅展示true
非投资建议true
预注册文档(sha256 钉定)c0ecaadfc317278b1ca514528a22f4415e17727be9da0ccf48964114a30fad51预注册原文(GitHub)

数据输入

字段值
fundamentals sha256df1f27cecc20acb6f2b5e2d75e88618733e42cec07c48187825a53330151ad7b
prices sha256c86551c35e3503b02565eada4c8a5e343ad40e008e26f47b2b99e9c36876dc46
membership sha25659374fda51c515eadd4cb7565c7f896abe54d95fae4a79476b1e157434eb2bf3
期末滞后(申报后月数)10-K: 6 · 10-Q: 4
特征列9 · mktcap, pb_ratio, roa, fund_assets, fund_revenue, fund_net_income, fund_equity, fund_shares_out, fund_long_term_debt
宇宙语义Universe membership is point-in-time (resolved as-of each historical panel date, never a today snapshot); fundamentals enter via filing dates. The input sha256 pins identify the exact frozen data artifacts behind the feature table.

模型

字段值
学习器lightgbm 4.7.0
库版本lightgbm 4.7.0 · purgedcv 0.1.2 · arch 8.0.0
冻结超参objective=regression · n_estimators=500 · learning_rate=0.05 · num_leaves=31 · min_child_samples=20 · reg_lambda=1 · feature_fraction=0.8 · bagging_fraction=0.8 · bagging_freq=1 · n_jobs=1 · random_state=0 · bagging_seed=0 · feature_fraction_seed=0 · drop_seed=0 · verbose=-1
确定性种子random_state=0 · bagging_seed=0 · feature_fraction_seed=0 · drop_seed=0 · n_jobs=1

评估协议

字段值
CV 方案5-fold PurgedGroupKFold (group=month, embargo=21 sessions) over the 2016+ resolvable window; the explicit 2011-2016/2017-2026 split in pre-reg §8.0 is superseded because the resolvable universe is 2016+
horizon21
n_splits5
embargo sessions21

结果(头条实值逐字嵌入)

字段值
combined rank-IC-0.0088
置信区间(95%)[-0.0336, 0.0159]
p(HAC)0.484
实现月数71
裁决NULL
SESOI0.01
结果 config sige14b9d44
ledger 行#49

治理 · 账本链

字段值
冻结行(config_committed)#27 · 2026-07-28T04:15:16+00:00
冻结行(config_committed) · sha256818c94fd55a03966d05e8452febbb8772db06145c171116ce8440f18003c9abb
结果行(confirmatory:first)#28
契约The config_committed row precedes the result row sharing its sha256 — the anti-leakage ordering, verifiable against runs/ledger.jsonl.
uv.lock(冻结时)e045a023c5ea8d9a91ef41f183dc95447f61788d79c5fa974229c7b32d5b34a0
uv.lock(当前)5c5cff1c84ece3e556fce275191377bca3eb2f84d97326527e3439a7e9cbc845
uv.lock 一致falseDependency lock drifted since the freeze (later repo lanes added dependencies); the frozen training environment stays pinned by uv_lock_sha256 and versions. The current lock hash is embedded beside it for honest comparison.
账本审计(/discipline)账本审计(/discipline)

溯源

字段值
生成器scripts/export_terminal_data.py::export_model_card (model_card_version v1)
再生成命令uv run python -c "import sys; sys.path.insert(0,'scripts'); import export_terminal_data as m; m._safe_export('model_card', m.export_model_card)"
字节稳定Zero-clock panel: no snapshot timestamp and no wall-clock reads; every value is machine-read verbatim from the pinned artifacts, so consecutive exports of unchanged artifacts are byte-identical. Output is LF.
源(S 编号)S1 frozen artifact · runs/results/17245a75d2d4cd17c68f36a9d0f4b4f7f3baf1db31b33b87be79e6e4a11400da/config.json — sha256:21da28dd91747ae5bc242867ac13850569c3109a59123c5f5c52cf21e6d363e7 (file bytes)S2 frozen artifact · runs/results/17245a75d2d4cd17c68f36a9d0f4b4f7f3baf1db31b33b87be79e6e4a11400da/meta.json — sha256:4eb28d84645378953392bcc51ffa9a25f0ed74dc69cbddf32c39f3ec9093f088 (file bytes)S3 ledger row · runs/ledger.jsonl line 27 — sha256:818c94fd55a03966d05e8452febbb8772db06145c171116ce8440f18003c9abb (line text)S4 ledger row · runs/ledger.jsonl line 28 — sha256:d8405c22044b2470928f901801ab27ef2d7e401cb7f5d676f0b1bfa64683b5ed (line text)S5 panel · web/src/data/aionis/metrics.json — verbatim value embed (results section); the source file carries a snapshot_ts clock, so no file hash is pinnedS6 repo doc · docs/track-c-preregistration.md — sha256:c0ecaadfc317278b1ca514528a22f4415e17727be9da0ccf48964114a30fad51 (LF bytes)S7 repo file · uv.lock — sha256:5c5cff1c84ece3e556fce275191377bca3eb2f84d97326527e3439a7e9cbc845 (LF bytes)

展示层档案——非投资建议,不构成任何交易信号;冻结面只在冻结工件变化时更新。

模型清单
SR 11-7 式模型清单:账本内每条 confirmatory:first 运行与冻结结果目录对账(冻结→结果行链 · H6 · 差分),config-only 提交如实列示。

覆盖账本行截至: 2026-08-08

阶段config_sig(前 12)冻结 ts(ledger)H6差分 mean_diff [95% CI]零假设本地目录
B17245a75d2d4#27 · 2026-07-28T04:15:16+00:00PASS-0.0008 [-0.010573, 0.008972]未知v0.1.0
Ca7fdb48f5942#29 · 2026-07-28T12:01:16+00:00PASS-0.006487 [-0.019531, 0.006556]未知v0.1.0
Dd31580630ff3#33 · 2026-07-29T02:31:16+00:00PASS-0.00298 [-0.013714, 0.007755]未知v0.1.0
E1ef321e9ee808#35 · 2026-07-29T03:29:39+00:00PASS-0.002793 [-0.011457, 0.005871]未知v0.1.0
track_ce14b9d445411#48 · 2026-08-05T04:19:50.785179+00:00——未知无目录(诚实未知)
config-only 提交: 11

已提交账本、本地无结果目录——可能被后续配置取代或目录未保留(原因不猜测);不逐行展开,明细见 model_inventory.json。

建模决策谱系

track_c账本序演化
#46ΔIC —#47ΔIC —#48ΔIC -0.0088 · p 0.484
track_adaptive显式取代
#51取代理由(ledger amendment 原文)amend1: WEEKLY cadence + truly-frozen baseline. Supersedes #51 (monthly-A, redundant with Track C).ΔIC —#52取代理由(ledger amendment 原文)amend2: n_estimators 500→100 (feasibility ~5min vs ~27min; both arms; comparison self-contained, valid at any n_estimators). Supersedes #52.ΔIC —#53ΔIC —

显式取代 = ledger amendment 原文声明(Supersedes #NN),取代理由挂在被取代行下;账本序演化 = 按账本序推断的排列,非显式取代声明。emerald 行 = 该配置已有 confirmatory 结果或本地冻结目录。

清单为展示层治理记录——runs/ 只读;无本地目录=诚实未知而非不存在;非投资建议。

供应商知识年限
LLM 知识截止的冻结经验边界与漂移告警——防 LLM 回忆泄漏的年限纪律(探针工件可核验)。
冻结截止
2023-03-10
出处
empirical-probe-v1 (dual-blind dated-event ladder; NOT a vendor declaration) — frozen in config/e3_live_contracts.yaml
探针边界
known ≤ 2023-03-10 · unknown ≥ 2024-11-06
最近漂移检查
stable

经验探针边界不是厂商声明;安全向漂移(模型学得更新)也必须走显式契约修订,绝不静默改值。钉住的 LLM 在冻结样本外管线中从不产生信号。

方法 + 边界

Leakage-safe model-drift monitor (display-only). Population Stability Index (PSI) compares the recent realized score distribution to the historical one; rolling cross-sectional rank-IC (recent mean vs full mean) is reported for context. All metrics use only realized (score, forward_return) pairs strictly before the latest unread month. PSI<0.1 stable / 0.1–0.25 moderate / ≥0.25 significant. NOT a research claim and NOT a retrain trigger (rerun-to-significance is forbidden).

信息性显示——非研究 claim,不触发重训(rerun-to-significance 禁)。