A0889
Title: Weighted maximum likelihood for misspecified mixed causal non-causal autoregressive models
Authors: Gabriele Mingoli - Vrije Universiteit (Netherlands) [presenting]
Francisco Blasques - VU University Amsterdam (Netherlands)
Siem Jan Koopman - VU Amsterdam (The Netherlands)
Abstract: The aim is to introduce a novel weighted maximum likelihood estimation (WMLE) method aimed at improving the forecast accuracy of misspecified models. The approach is specifically designed for mixed causal non-causal autoregressive (MAR) models, which allow a stochastic process to depend on its future values through a lead polynomial, capturing locally explosive dynamics. MAR models are commonly used to describe and predict speculative bubbles in financial time series. However, they impose a rigid structure on all bubbles in a given dataset, implying that all bubbles exhibit the same dynamics, which may not always align with empirical observations. In reality, explosive episodes in time series often vary in size, duration, and growth rate, making the standard MAR specification restrictive. It is demonstrated that by applying different weights to different parts of the sample when estimating a MAR model, it is possible to construct an estimator that enhances forecasting performance compared to traditional maximum likelihood estimation (MLE). A simulation study confirms that the proposed weighting approach improves predictive accuracy under misspecification. Finally, an out-of-sample forecasting exercise on monthly crude oil prices shows that WMLE outperforms forecasts obtained using standard MLE.
