Walk-forward CV on the same 8 folds as the model comparison. Each
configuration caps the training windows with
make_data_op(max_train_rows=N), so only the amount of history used
for fitting changes; test dates are identical. Metrics are pooled over all
folds. TabICL uses eight estimators and its median prediction for MAPE.
"all" means no cap: each fold uses all data since 2021-03, about 26–47 months depending on the fold, so it has no single months value and is shown in the tables but not on the month axis. Training-history values are exact months (rows / 730.5 hours per month).
| Training history | MAPE 1h | MAPE 12h | MAPE 24h | MAPE average | Mean fit time / fold |
|---|---|---|---|---|---|
| 2.957 mo | 4.51% | 7.37% | 7.60% | 6.49% | 10.4 s |
| 11.992 mo | 1.27% | 2.38% | 2.42% | 2.02% | 21.6 s |
| 23.984 mo | 1.15% | 2.10% | 2.17% | 1.81% | 25.7 s |
| 35.975 mo | 1.11% | 1.95% | 2.01% | 1.69% | 30.9 s |
| all | 1.11% | 1.92% | 1.99% | 1.67% | 33.0 s |
| Training history | MAPE 1h | MAPE 12h | MAPE 24h | MAPE average | Total inference time |
|---|---|---|---|---|---|
| 2.957 mo | 2.92% | 6.05% | 6.18% | 5.05% | 63.0 s |
| 11.992 mo | 0.89% | 2.25% | 2.54% | 1.90% | 164.2 s |
| 23.984 mo | 0.83% | 2.04% | 2.34% | 1.74% | 307.7 s |
| 35.975 mo | 0.79% | 2.07% | 2.14% | 1.67% | 422.7 s |
| all | 0.79% | 2.06% | 2.12% | 1.65% | 463.1 s |
Plots show the last CV fold (2025-03-01 to 2025-05-31) so the curves stay readable; quality metrics above use all 8 folds.