In the COVID quarters every model's error is several times its error afterwards, and two of the three foundation models go from lowest on the chart to highest. Foundation models say nothing about why demand moves.
What this shows: The chart plots Post-COVID (expanding 1-step) and COVID 2020 Q1 to 2021 Q2 (expanding 1-step) on a vertical axis of RMSE (lower is better).
Conditions and method
Status: descriptive (the models compared do not see the same information: the traditional forecasters condition on realised same-quarter macro inputs and the foundation models do not).
ARDL, Neural Ensemble, NeuralProphet and State Panel FE are not scored on this chart, because they have no expanding-window evaluation. The title counts only the models that are scored. On each model’s plain RMSE over its own evaluation window the traditional models score lower, but the windows differ (the foundation models’ includes the COVID years), so that ordering is not like for like either.
RMSE is the root mean squared error of forecasts made for the next quarter.
The models are in ascending order of post-COVID error.