Learning curves for weak signals

Training error is low but chronological validation error is high. What experiments help distinguish excessive flexibility from a changing data distribution?

Compare learning curves, stronger regularisation, simpler models and rolling evaluation windows. If more comparable training data and lower flexibility close the gap, variance is a plausible explanation. If errors track regime changes or feature shifts, investigate distribution change and stale relationships. These explanations can coexist. Preserve an outer evaluation window while running these diagnostics so experiments do not manufacture apparent validation progress.