Before the pandemic the leading models were close together. Afterwards the seasonal model predicts best and the error-correction model worst of those scored here; each sees the quarter's actual economic data.
What this shows: The chart plots Pre-COVID holdout (2017–2019) and Post-COVID generalisation (2021 Q3 onward, expanding window) on a vertical axis of RMSE (lower is better).
Conditions and method
Status: descriptive (accuracy is conditional on feeding models the realised same-quarter macro inputs, not a real-time forecast test).
All models on this chart condition on realised same-quarter macro data: each is given the actual GDP, CPI and oil price of the quarter it forecasts, so these are not the errors of a forecast made in real time. The foundation models on the cross-family chart see macro data only up to the quarter before, so a comparison with them favours the models shown here. The expanding-window ElasticNet predictions are capped at three standard deviations from the training mean.
ARDL has no post-COVID score on this chart, so the post-COVID ranking covers the other models only. The ARDL plotted here is a forecasting version fitted on data to the end of 2019. It is a different model from the one behind the site’s headline elasticity, and that model is not scored on this chart.
The ‘static’ column mixes four ways of forecasting that cannot be compared, so its ranking of the models is not like for like. The expanding-window column is the matched comparison.