The Focused Information Criterion for Stochastic Model Selection Problems Using $M$-Estimators

TL;DR

This paper extends the focused information criterion (FIC) to general M-estimators for stochastic model selection, providing asymptotic theory and demonstrating improved performance over AIC and BIC in simulation studies.

Contribution

It develops the FIC framework for M-estimators in stochastic processes, including asymptotic theory and practical applications in spatial models.

Findings

FIC outperforms AIC in correct model selection.

FIC performs comparably to BIC in simulations.

Applications include autoregressive models and density selection.

Abstract

Claeskens and Hjort (2003) constructed the focused information criterion (FIC) and developed frequentist model averaging methods using maximum likelihood estimators assuming the observations to be independent and identically distributed. Towards the immediate extensions and generalizations of these results, the present article is aimed at providing the focused model selection and model averaging methods using general maximum likelihood type estimators, popularly known as $M$-estimators. The necessary asymptotic theory is derived in a setup of stationary and strong mixing stochastic processes employing von Mises functional calculus of empirical processes and Le Cam's contiguity lemmas. We illustrate the proposed focused stochastic modeling methods using three well-known spacial cases of $M$-estimators, namely, conditional maximum likelihood estimators, conditional least square estimators and estimators based on method of moments. For the sake of simulation exercises, we consider two simple applications of FIC. The first application discusses the simultaneous selection of order of autoregression and symmetry of innovations in asymmetric Laplace autoregressive models. The second application demonstrates the FIC based choice between general scale-shape Gamma density and exponential density with shape being unity. We observe that in terms of the correct selections, FIC outperforms classical Akaike's information criterion AIC and performs at par with Bayesian information criterion BIC.