Evaluate trained constraints at portfolio scale

Determine whether trained constraints improve on manual selection across the full portfolio of Schedule P company-line datasets.

Background

The paper demonstrates its supervised-learning framework on a single real Schedule P triangle. The training loop selects experience-related hyperparameters, including recency, reference-pattern, and smoothness weights, using held-out calendar diagonals. The example shows that these trained constraints can produce a plausible reserve estimate, but the dataset was selected for illustrative purposes rather than as part of a portfolio-wide evaluation.

Consequently, the paper does not establish whether the trained approach consistently outperforms manual pattern selection across the broader population of Schedule P company-line datasets. A portfolio-scale empirical comparison is therefore left unresolved.