All the CV test folds are pooled. Models: chronos2_ctx4000, hgb_quantile_q7, hgb_quantile_recal_q7, kumo_tabular, tabicl_forecast_ctx1024_n4, tabicl_n8_full. Nominal interval 0.05–0.95 (nominal coverage 0.90).
For each nominal level τ, the fraction of observations at or below the τ-quantile forecast. Dashed line = perfect calibration. This is the marginal check; a model can be on the diagonal on average while being miscalibrated in particular situations.
Test cases split into 10 equal-count bins by each model's predicted median; observed coverage of the 0.05–0.95 interval in each bin. Below the dashed nominal line means the interval is too narrow at that load, above means too wide.
Equal-count bins of each model's own predictions; x = mean prediction, y = mean prediction + τ-quantile of the residuals in the bin. A perfectly calibrated model lies on the dashed 45 degree diagonal (1:1 axes). Use the dropdown to change the horizon.
Same check with the conditional τ-quantile estimated by isotonic regression (no bin width to choose). A perfectly calibrated model lies on the dashed 45 degree diagonal (1:1 axes).