Sujet 3 : Forecasting with model uncertainty: Theory and applications


Forecasting with model uncertainty: Theory and applications

A good forecasting model is of crucial importance to governments and business leaders for making policy decisions. Often there are a number of predictors for a variable of interest. Instead of focusing on the selection of the best forecasting model, this work proposes new methods for optimal forecast combination to deal with model uncertainty surrounding any econometric estimation aiming to provide predictions of economic variables. Optimal forecast combination or model averaging have been used in various settings for inference, prediction or policy analysis to address model uncertainty. For instance, recent literature investigated the issue of optimal forecast in policy evaluation, macroeconomic forecasting, finance or analysis of economic growth, among others. Existing methods delivering combinations such as equal weighting, BIC or AIC selection, BIC or AIC weighing, Bates-Granger combination, predictive least squares, Mallows Model Averaging or jacknife model averaging do not successfully deal with standard features of macroeconomic or financial time-series such as serial dependence or time-varying volatility, among other issues. We develop new algorithms based on Bayesian or frequentist schemes to fill this gap and show their accuracy in simulations and empirical applications. We also include economic measures of forecasting performance such as utility-, asymmetric- or regret-based measures to gauge the performance of our new procedures in various environments. Therefore, our analysis emphasizes the interplay between (1) the economic-based aspect of the performance measure of predictive content and (2) the statistical accuracy of econometric models aiming at delivering forecasts of economic variables.


Pre-requisite: Économétrie; Méthodes numériques, Méthodes Bayésiennes (non indispensables)


Laboratoire d’accueil : LEMNA (Université de Nantes)


Return >>  Offres de thèses