Figure 2.

Comparing the best models for Bangkok, by model type. Focusing only on the best performing model for each model type and each time horizon, we show the relative mean absolute error (left panel) and the mean absolute error (right panel). On the left, the baseline of comparison is the traditional AR(1) model and the y-axis can be interpreted as the improvement over this baseline—i.e., a value of .9 indicates a 10% improvement. We show that both the Plain SARIMA (red) and CDR SARIMA (green) models perform better than the LASSO model at earlier forecasting horizons but perform worse at later horizons.