加载中...
本报告提出一种漏斗/泄露防护的两阶段机器学习框架:阶段一用XGBoost二分类器识别是否发生发票稀释,阶段二用多种回归器(XGBoost/RandomForest/MLP/FasterKAN)及加权集成估计稀释金额,在7个滚动时间窗的验证中加权集成实现均值RMSE≈1215.8±99.7、WMAPE≈16.82%(MAPE≈17%–19%);分类器ROC-AUC稳定在0.9167–0.9222区间(最近窗0.9222),并采用仅使用先前发票的“无泄露”时间序列特征与宏观变量增强输入特征(137个特征,样本≈4.84M),为实时动态授信与风险定价提供可部署的预测模块
| Calendar window | ROC-AUC | Avg. precision | Recall (Non-diluted) | Recall (Diluted) |
|---|---|---|---|---|
| 2015-2019 | 0.9218 | 0.8457 | 79.01% | 94.30% |
| 2016-2020 | 0.9222 | 0.8498 | 88.05% | 77.45% |
| 2017-2021 | 0.9202 | 0.8428 | 82.72% | 88.61% |
| 2018-2022 | 0.9167 | 0.8493 | 84.43% | 83.14% |
| 2019-2023 | 0.9200 | 0.8446 | 88.64% | 74.20% |
| 2020-2024 | 0.9204 | 0.8361 | 80.95% | 90.31% |
| 2021-2025 | 0.9222 | 0.8369 | 88.68% | 75.67% |
| Model | R² | MAE | RMSE | MAPE (%) | WMAPE (%) |
|---|---|---|---|---|---|
| XGBoost | 0.8433 ±0.0212 | 297.0914 ±9.9873 | 1372.5669 ±85.9816 | 18.0758 ±0.7138 | 24.5048 ±1.0166 |
| RandomForest | 0.8585 ±0.0233 | 289.1397 ±8.3774 | 1302.6359 ±101.0308 | 17.7901 ±0.6583 | 23.8512 ±0.9621 |
| FasterKAN | 0.8481 ±0.0323 | 306.6415 ±14.8216 | 1347.7169 ±143.2914 | 19.1279 ±0.4736 | 25.2876 ±1.2554 |
| MLP | 0.8632 ±0.0234 | 285.0667 ±0.0234 | 1281.3949 ±110.5764 | 17.8387 ±0.3303 | 23.5136 ±0.9712 |
| Ensemble (AVG) | 0.8712 ±0.0200 | 283.7568 ±9.2897 | 1243.5975 ±90.8746 | 17.2399 ±0.4146 | 16.9059 ±0.5411 |
| Ensemble (WGT) | 0.8768 ±0.0209 | 282.3884 ±8.7000 | 1215.8378 ±99.6911 | 17.5501 ±0.3922 | 16.8238 ±0.4842 |
元数据与概览(引言与报告概览)
逐节深度解读
图表与表格的逐一深度解读(关键环节)
图像/可视化(每幅图单独描述、解读并链接文本)
估值分析(本研究中不适用)
风险因素评估(报告识别与隐含风险)
批判性视角与细微差别(审慎点评)
结论性综合
参考与后续建议(基于报告内容)
注:本文中对报告事实性陈述均直接依据原文相应页码加以标注以便溯源和复核;同时已将报告中所有主要表格与图像的关键信息逐一解读并以原始图片路径嵌入呈现以便对照验证(见上文各图像引用) .
2602.15248.pdf