python 3.12.13, scikit-learn 1.9.1, pandas 3.0.6
rows: dev 15867, holdout 1344; forecast origins 1321, horizon 1..24

MAE by horizon (bikes per hour, lower is better), default HistGradientBoostingRegressor
  strategy                                    h=1    h=2    h=3    h=6   h=12   h=24  mean 1-24
  one-step, real lag_1 (not a forecast)      24.5   24.5   24.5   24.4   24.5   24.6       24.5
  recursive                                  24.5   32.2   37.9   46.1   47.8   49.4       45.7
  direct                                     24.5   36.5   50.5   50.2   53.5   50.4       49.4
  lag_24 + lag_168 only                      49.7   49.7   49.7   49.6   49.9   50.4       50.0
  no lags                                    53.2   53.2   53.2   53.2   53.9   54.6       53.9

  direct, MAE at every horizon 1..24: 24.5 36.5 50.5 50.0 52.3 50.2 49.1 51.0 49.3 51.9 45.5 53.5 49.8 52.6 49.1 49.0 56.0 52.8 52.5 51.5 52.1 55.8 49.1 50.4

Same, early_stopping=False (every model trained on all dev rows)
  strategy                                    h=1    h=2    h=3    h=6   h=12   h=24  mean 1-24
  one-step, real lag_1 (not a forecast)      23.8   23.8   23.8   23.8   23.9   24.1       24.0
  recursive                                  23.8   31.9   37.6   45.4   47.2   47.5       44.8
  direct                                     23.8   35.4   45.8   51.8   47.8   46.4       47.0
  lag_24 + lag_168 only                      45.6   45.6   45.6   45.5   45.9   46.4       45.9
  no lags                                    50.1   50.1   50.1   50.1   50.7   51.4       50.7

  direct, MAE at every horizon 1..24: 23.8 35.4 45.8 45.9 48.9 51.8 52.0 47.7 47.0 49.8 47.2 47.8 48.0 48.4 50.5 52.4 46.3 51.0 47.4 48.5 49.4 47.2 49.3 46.4

Training cost on the dev set (default settings)
  one-step model            1 fit(s),   0.3 s
  direct (24 models)       24 fit(s),   7.2 s
  lag_24 + lag_168 model    1 fit(s),   0.3 s