Pareto plots: oracle vs forecast weather

Times are summed over the 8 CV folds (fit + predict/score). The new forecast-weather HGB runs use cpu-1 with all 8 cores sequentially; the oracle HGB timings are from the older run with OMP_NUM_THREADS=2 and concurrent HGB jobs, so HGB oracle vs forecast times are not directly comparable. Accuracy deltas are unaffected. GPU runs use gpu-2; CPU/GPU hardware differs, so times are indicative rather than a controlled benchmark. Table values are means across the 8 folds. Lower is better on both axes. KumoForecast is absent from the pinball chart because it has no quantiles.

Paired differences vs source-specific baseline

Each model is compared with its own weather-source HGB baseline (HGB mean for MAPE, HGB q7 for pinball loss). 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 per-source Pareto frontiers.

Absolute accuracy vs fit + predict time

Means across the 8 CV folds, shown without error bars. Dashed colored lines mark the per-source Pareto frontiers.