Times are summed over the 8 CV folds (fit + predict/score). They come from per-fold timings in the scores CSV when available, otherwise from the run log. Hardware and parallel load differ across runs (CPU vs GPU), so times are indicative rather than a controlled benchmark. The horizontal axis is log-scaled. MAPE uses the median prediction. Pinball loss is averaged over the 5%, 50% and 95% quantiles and all horizons; models without quantiles are omitted from the pinball figure. Lower is better on both axes. Table values are means across the 8 folds.
| model | fit + predict (s) | MAPE (%) | pinball loss (MW) |
|---|---|---|---|
| hgb_mean_fast | 55 | 1.801 | - |
| hgb_mean | 252 | 1.784 | - |
| tabicl_n8_full | 523 | 1.704 | 225.95 |
| kumo_forecast | 615 | 3.777 | - |
| kumo_tabular | 1,273 | 1.405 | 190.70 |
| hgb_quantile_q7 | 2,154 | 1.753 | 244.21 |
| chronos2_ctx4000 | 3,294 | 1.706 | 219.23 |
| hgb_quantile_recal_q7 | 8,921 | 1.844 | 244.84 |
| tabicl_forecast_ctx1024_n4 | 12,051 | 2.141 | 280.12 |
MAPE is relative to hgb_mean;
pinball loss is relative to hgb_quantile_q7.
The baseline is shown at zero difference. Large markers: pooled mean difference over all test rows, with a 95% moving-block bootstrap interval (14-day blocks over daily differences; conditional on the fitted models, it ignores training randomness). Small dots: mean difference in each of the 8 CV folds (one dot per fold; hover shows the fold). Folds are not independent, so no t-interval across folds is shown. Dashed colored lines
mark the Pareto frontiers.
Means across the 8 CV folds, shown without error bars. Dashed colored lines mark the Pareto frontiers.