This paper compares feed‑forward neural networks (FNN) trained on summarised transactional data with recurrent architectures (LSTM) trained on full payment and case estimate time series, using 50 complex SPLICE‑simulated datasets; it finds that including case estimate histories materially improves predictive performance, while adding recurrent memory (LSTM) provides only marginal improvements over a well‑designed FNN in aggregate and often deteriorates individual‑level metrics, and presents a standardised evaluation framework for insurers to assess case estimate value .
Table 1: Reserve error (OCLerr), MALE and MSLE (mean across 50 datasets) for LSTM+, FNN+ and Case Estimates
| Model | OCLerr (%) mean (sd) | MALE mean (sd) | MSLE mean (sd) |
|---|---|---|---|
| Case Estimates | -33.6 (0.8) | 0.828 (0.017) | 2.051 (0.066) |
| LSTM+ | -6.0 (5.7) | 0.629 (0.042) | 0.934 (0.218) |
| FNN+ | -7.7 (6.9) | 0.550 (0.036) | 0.633 (0.068) |
Table 3: FNN+ (with case estimates) vs FNN (without)
| Model | OCLerr (%) mean (sd) | MALE mean (sd) | MSLE mean (sd) |
|---|---|---|---|
| Case Estimates | -33.6 (0.8) | 0.828 (0.017) | 2.051 (0.066) |
| FNN+ | -7.7 (6.9) | 0.550 (0.036) | 0.633 (0.068) |
| FNN | -20.5 (6.0) | 0.773 (0.060) | 1.182 (0.254) |
元数据与概览(引言与报告概览)
逐节深度解读 — 结构与关键要点
图表与图片的逐一深度解读(关键表格与图像)
训练、数据与评估细节剖析
估值、超参数与实现细节(Appendix D 与训练方法)
风险因素评估与作者列举的局限
批判性视角与细微之处(基于报告内容的审慎评价)
结论性综合(要点提炼)
若需我进一步逐图逐表地对每个箱线图与分季度图做数值级(例如均值、中位数、上下四分位与异常值)解读并把对应图片逐一嵌入报告位置(便于直接复制到内部报告或讲稿中),我可以继续展开并在每个图注后追加对应的原文页码溯源引用。
2601.05274.pdf