From Chain-Ladder to Individual Claims Reserving

Battle Count: 72

This paper is squarely in the domain of actuarial science and insurance claims reserving. While the multi-period forecasting framework and recursive estimation concepts have some structural parallels to financial time series prediction, the paper has no direct application to quantitative trading, portfolio management, or market prediction. The methodology is specific to insurance liability estimation.

The core FNN architecture is simple (3 hidden layers, 606-646 weights), but the overall implementation complexity is moderate-to-high due to: (1) the recursive estimation procedure requiring careful cohort construction at each step, (2) bias control via balance correction, (3) data pre-processing with log-transforms and standardization, (4) ensembling over 10 network fits, (5) the need to fit separate networks per development period, and (6) the requirement for consistent claims cohort selection. The conceptual framework (PtU factors) is elegant but the practical implementation requires careful attention to the recursive structure.

The paper provides detailed FNN architecture specifications (Table 3), hyperparameters (Table 4), data pre-processing steps, and the full recursive algorithm. However, the real insurance datasets (accident and liability) are not publicly available. The methodology is clearly described and the proof of Proposition 2.2 is provided in the appendix. The Lorenz-Schmidt reference (2016) is cited as having established the main result earlier.