imputation_method='drift'. ADF test confirms the imputed series is non-stationary (p=0.95), cleanly capturing the trend+seasonality the models need to learn.
| Model | Type | MAPE | SMAPE | RMSE | MAE | R² | MASE |
|---|---|---|---|---|---|---|---|
| Exponential Smoothing ★ #1 | Statistical | 5.43% | 5.48% | 81.60 | 68.37 | 0.8293 | 0.8150 |
| Extra Trees CDS+DT #2 | ML | 6.01% | 5.94% | 90.86 | 76.77 | 0.7638 | 0.9256 |
| OMP CDS+DT #3 | ML | 6.27% | 6.14% | 87.85 | 77.91 | 0.8346 | 0.9179 |
| Gradient Boosting CDS+DT | ML | 6.52% | 6.45% | 97.02 | 83.88 | 0.7165 | 1.0126 |
| Random Forest CDS+DT | ML | 6.74% | 6.63% | 101.10 | 85.32 | 0.7094 | 1.0271 |
| Decision Tree CDS+DT | ML | 6.41% | 6.64% | 112.45 | 90.06 | 0.7111 | 1.0579 |
| K Neighbors CDS+DT | ML | 7.16% | 6.95% | 107.66 | 91.61 | 0.6763 | 1.0946 |
| LightGBM CDS+DT | ML | 7.21% | 7.01% | 105.71 | 90.86 | 0.6397 | 1.1015 |
| AdaBoost CDS+DT | ML | 7.97% | 7.74% | 114.70 | 99.00 | 0.6009 | 1.1980 |
| ARIMA | Statistical | 8.65% | 8.97% | 118.88 | 109.81 | 0.2589 | 1.3680 |
| Theta Forecaster | Statistical | 9.04% | 9.76% | 118.84 | 111.68 | 0.3182 | 1.3855 |
| Seasonal Naive | Baseline | 9.31% | 9.51% | 138.90 | 121.54 | 0.3747 | 1.4778 |
| Huber CDS+DT | ML | 9.33% | 9.87% | 142.19 | 120.82 | 0.1714 | 1.4851 |
Strong seasonal structure survives imputation
Drift interpolation preserves the pre/post-pandemic seasonal peaks. Exp. Smoothing's additive Holt-Winters structure matches this pattern directly — no feature engineering needed.
ML models overfit to reconstructed data
Extra Trees and OMP apply Conditional Deseasonalize & Detrending (CDS+DT) before fitting, but their tree-based residual fitting struggles with the imputed plateau — explaining the higher MAPE vs. a pure statistical model.
OMP wins on R² — but not on MAPE
Orthogonal Matching Pursuit achieves the highest R² (0.8346 vs 0.8293 for Exp. Smooth) because it better captures variance across the full test set. However, its MAPE is higher (6.27%), meaning its errors are proportionally larger on individual months — the deciding criterion here.
Blended ensemble: conservative but stable
The blend of the three models reduces peak forecasts (e.g. Jul 2026: 1,906k vs 1,952k for Exp. Smooth alone) but also reduces the chance of overestimation. For operational planning, the blend is the safest recommendation.