Regularization: Weight Decay, Dropout, Early Stopping, and Validation Discipline

A volatility-regime classifier that drives training cross-entropy to near zero on last year's books usually fails on next month's regime shift: capacity plus unconstrained fitting invents fragile co-adaptations among correlated features. This note separates four levers that keep a network honest — $\ell_2$ / weight decay, dropout, early stopping, and a held-out validation protocol — so you can constrain what is fitted and when fitting stops without confusing them with the optimiser update.