项目: interval-financial-risk | 随机种子: 202608
合成数据演示 · 结果用于检验研究流程，不代表真实企业风险识别能力。
比较点特征与过去四个季度的分布特征；同一公司的未来季度进入测试集。报告期是合成季度标记，尚未证明真实财报发布日期可得性。
{"source": "synthetic", "rows": 4800, "companies": 300, "quarters": 16, "positive_rate": 0.12, "date_start": "2019-03-31", "date_end": "2022-12-31", "data_sha256": "75acf0ca44c06957ae6d62e2225280d2626b1304bb00d67bd36d92fd63dbd336"}
表格为各测试折指标的等权均值 ± 折间标准差；ROC/PR 图将测试预测合并，口径不同。标准差不是置信区间。分类阈值固定 0.5；PR-AUC 使用 average precision。
诊断模型由验证集平均 AUC 选择：Logistic Baseline。消融与置换重要性在最后测试折评估，属于探索性诊断。
按 report_date 扩展训练窗口；预处理仅拟合训练集。验证窗口用于模型选择，测试窗口不参与该选择。
| fold | train_start | train_end | train_n | train_positives | validation_start | validation_end | validation_n | validation_positives | test_start | test_end | test_n | test_positives |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 2019-03-31 | 2020-12-31 | 2400 | 193 | 2021-03-31 | 2021-06-30 | 600 | 94 | 2021-09-30 | 2021-12-31 | 600 | 84 |
| 2 | 2019-03-31 | 2021-06-30 | 3000 | 287 | 2021-09-30 | 2021-12-31 | 600 | 84 | 2022-03-31 | 2022-06-30 | 600 | 103 |
| 3 | 2019-03-31 | 2021-12-31 | 3600 | 371 | 2022-03-31 | 2022-06-30 | 600 | 103 | 2022-09-30 | 2022-12-31 | 600 | 102 |
| model | auc | pr_auc | brier | f1 | precision | recall | calibration_error | auc_std | pr_auc_std | brier_std | f1_std | precision_std | recall_std | calibration_error_std |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Logistic Baseline | 0.724 | 0.344 | 0.203 | 0.402 | 0.309 | 0.591 | 0.281 | 0.024 | 0.055 | 0.010 | 0.037 | 0.048 | 0.030 | 0.026 |
| Logistic + Interval Features | 0.686 | 0.310 | 0.218 | 0.363 | 0.268 | 0.580 | 0.282 | 0.034 | 0.057 | 0.020 | 0.045 | 0.048 | 0.029 | 0.037 |
| Elastic Net + Interval Features | 0.686 | 0.308 | 0.218 | 0.362 | 0.268 | 0.576 | 0.281 | 0.035 | 0.061 | 0.020 | 0.047 | 0.050 | 0.026 | 0.036 |
| XGBoost | 0.678 | 0.302 | 0.147 | 0.340 | 0.382 | 0.306 | 0.106 | 0.036 | 0.061 | 0.002 | 0.065 | 0.079 | 0.058 | 0.013 |
| LightGBM | 0.669 | 0.325 | 0.154 | 0.343 | 0.333 | 0.354 | 0.125 | 0.035 | 0.055 | 0.003 | 0.045 | 0.057 | 0.032 | 0.016 |
预先固定为 Logistic + Interval Features，各消融变体重新拟合，使用同一最后测试折；这样即使点模型获胜，移除区间特征也不会变成空操作。
名义覆盖 90% · 经验覆盖 92.2% · 平均宽度 0.965
覆盖对象是二元标签 y，不是未知真实风险概率。时间漂移与公司内相关性违反经典可交换性假设；只报告本次经验覆盖。区间很宽也可能得到高覆盖，应同时阅读宽度。
预先固定模型：Logistic + Interval Features；不按测试或校准得分挑选模型。
{"train": {"start": "2019-03-31", "end": "2021-03-31", "rows": 2700}, "calibration": {"start": "2021-06-30", "end": "2021-12-31", "rows": 900}, "test": {"start": "2022-03-31", "end": "2022-12-31", "rows": 1200}}
仅控制行业；随机交叉拟合未按公司分组，OLS 标准误未做公司聚类修正。系数不构成因果效应或稳健统计显著性证据。
生成命令：python scripts/run_experiment.py --config config/config.yaml。随附 run_manifest.json 记录依赖、数据摘要与源码哈希，predictions.csv 支持独立复算。
{
"project": {
"name": "interval-financial-risk",
"seed": 202608,
"n_jobs": 1
},
"data": {
"use_synthetic": true,
"synthetic": {
"n_companies": 300,
"n_quarters": 16,
"start_date": "2019-01-01",
"freq": "QE",
"risk_rate": 0.12
},
"raw_path": "data/raw",
"processed_path": "data/processed"
},
"features": {
"point_features": [
"revenue_growth",
"profit_margin",
"operating_cash_flow",
"volatility"
],
"interval_features": [
"revenue_growth",
"profit_margin",
"operating_cash_flow",
"volatility"
],
"interval_window": 4,
"interval_stats": [
"mean",
"std",
"min",
"max",
"q25",
"q50",
"q75",
"skew",
"kurt"
],
"use_interval_width": true,
"impute_strategy": "median"
},
"models": {
"logistic_baseline": {
"name": "Logistic Baseline",
"model_type": "logistic_regression",
"params": {
"max_iter": 1000,
"class_weight": "balanced"
},
"feature_set": "point_only",
"optimize": false,
"optuna_trials": 50
},
"logistic_interval": {
"name": "Logistic + Interval Features",
"model_type": "logistic_regression",
"params": {
"max_iter": 1000,
"class_weight": "balanced"
},
"feature_set": "point_and_interval",
"optimize": false,
"optuna_trials": 50
},
"elastic_net": {
"name": "Elastic Net + Interval Features",
"model_type": "elastic_net",
"params": {
"max_iter": 5000,
"class_weight": "balanced",
"l1_ratio": 0.5,
"C": 1.0
},
"feature_set": "point_and_interval",
"optimize": false,
"optuna_trials": 50
},
"xgboost": {
"name": "XGBoost",
"model_type": "xgboost",
"params": {
"n_estimators": 200,
"max_depth": 4,
"learning_rate": 0.05,
"subsample": 0.8,
"colsample_bytree": 0.8,
"eval_metric": "logloss"
},
"feature_set": "point_and_interval",
"optimize": false,
"optuna_trials": 50
},
"lightgbm": {
"name": "LightGBM",
"model_type": "lightgbm",
"params": {
"n_estimators": 200,
"max_depth": 4,
"learning_rate": 0.05,
"subsample": 0.8,
"colsample_bytree": 0.8,
"verbose": -1
},
"feature_set": "point_and_interval",
"optimize": false,
"optuna_trials": 50
}
},
"validation": {
"method": "rolling_window",
"initial_train_quarters": 8,
"test_quarters": 2,
"step_quarters": 2,
"min_train_samples": 200
},
"evaluation": {
"metrics": [
"auc",
"pr_auc",
"brier",
"f1",
"precision",
"recall"
],
"permutation_n_repeats": 20,
"ablation_groups": [
{
"name": "interval_stats",
"description": "移除所有区间统计特征",
"remove_patterns": [
"_mean",
"_cv",
"_std",
"_min",
"_max",
"_q25",
"_q50",
"_q75",
"_skew",
"_kurt",
"_width"
]
},
{
"name": "quantiles",
"description": "移除分位数特征",
"remove_patterns": [
"_q25",
"_q50",
"_q75"
]
},
{
"name": "width",
"description": "移除区间宽度特征",
"remove_patterns": [
"_width"
]
}
],
"calibration_bins": 10
},
"groups": {
"by_industry": true,
"by_time": true
},
"output": {
"figures_dir": "outputs/figures",
"reports_dir": "outputs/reports",
"models_dir": "outputs/models"
}
}