Using regression models with time-varying coefficients for short-term economic forecasting

This study describes three methods for estimating linear regression models with time-varying coefficients: piecewise regression, local regression, and regression with stochastic coefficients (state-space modeling). It also details their implementation in R using the tvCoef package. Through a comparative analysis of around thirty quarterly forecasting models, we show that the use of these methods, especially thanks to the state-space modeling, reduces forecast errors when breakpoints are present in the coefficients. Moreover, even when traditional tests conclude that the coefficients are stable, regression with stochastic coefficients can still help reduce forecast errors. However, uncertainties related to estimating certain hyperparameters can increase real-time forecast errors, especially for local regression. Thus, an economic analysis of estimated parameters remains essential.

This study is fully reproducible and all the codes used are available under
Ouvrir dans un nouvel onglethttps://github.com/InseeFrLab/DT-tvcoef.