Working Paper: CEPR ID: DP3312
Authors: Massimiliano Marcellino
Abstract: In this Paper we evaluate the relative performance of linear, non-linear and time-varying models for about 500 macroeconomic variables for the countries in the Euro area, using a real-time forecasting methodology. It turns out that linear models work well for about 35% of the series under analysis, time-varying models for another 35% and non-linear models for the remaining 30% of the series. The gains in forecasting accuracy from the choice of the best model can be substantial, in particular for longer forecast horizons. These results emerge from a detailed disaggregated analysis, while they are hidden when an average loss function is used. To explore in more detail the issue of parameter instability, we then apply a battery of tests, detecting non-constancy in about 20-30% of the time series. For these variables the forecasting performance of the time-varying and non-linear models further improves, with larger gains for a larger fraction of the series. Finally, we evaluate whether non-linear models perform better for three key macroeconomic variables: industrial production, inflation and unemployment. It turns out that this is often the case. Hence, overall, our results indicate that there is a substantial amount of instability and non-linearity in the EMU, and suggest that it can be worth going beyond linear models for several EMU macroeconomic variables.
Keywords: European Monetary Union; instability; nonlinear models; non-linearity; time-varying models
JEL Codes: C20; C53; E30
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| model type (C52) | forecasting accuracy (C53) |
| nonlinear models (C32) | forecasting performance for industrial production, inflation, and unemployment (E27) |
| linear models (C29) | forecasting performance (C53) |
| time-varying models (C32) | forecasting performance (C53) |