Contribution of non-linear models in forecasting macroeconomic aggregates: applications to the business cycle and monetary policy.

Summary This thesis proposes to explore the following paradox: although they significantly improve the goodness of fit of some macroeconomic aggregates, Markov regime-switching models do not seem to provide significantly more accurate forecasts than linear models. In a first chapter, we examine the robustness of this result for different specifications. Using a decomposition approach, we evaluate the contribution of the different sources of forecast error. Based on the decomposition results, the second chapter proposes a new two-step forecasting approach. In the first step, the effort is devoted to producing reliable forecasts of future plans that exploit the information provided by leading indicators. In a second step, we inject these new regime forecasts into the analytical formulations of predictors established in the previous chapter. Results on simulated and empirical data demonstrate the relevance of the proposed approach. In a third chapter, we examine possible asymmetries in the dynamics of U.S. inflation. The forecasting of inflation over the last twenty years once again highlights the contribution of the new two-step procedure. The fourth chapter is devoted to the study of asymmetries in the conduct of US monetary policy. By showing the limits of a linear representation, a non-linear rule is proposed to detect different monetary policy regimes. The analysis is enriched by estimating forward-looking rules using different measures of inflation expectations.