Two models on the same dates

How would you assess whether one forecasting model improves on another when both predict the same return observations?

Compare paired loss differences on common evaluation observations rather than treating model errors as independent samples. Examine their mean, economic magnitude and serial dependence; use an appropriate dependence-aware uncertainty estimate. Account for overlapping horizons and whether either model was selected using the same sample. Also inspect performance by period and asset group so a single mean does not hide concentrated failures.