Use time-series cross-validation (never peek at the future). - Compare the ensemble to each individual model. - Check calibration of the ensemble. - Analyze the diversity metrics to understand why the ensemble works.