A Monte Carlo Comparison of rMDL, gMDL, eMDL, nMDL, AIC and BIC Under the Effect of Stochastic Variance for Asymmetric Price Transmission Linear Models

Abstract

Stochastic variance or amount of noise has influenced Information-Theoretic Fit Criteria’s (ITFC) ability to recover the true DGP or select the optimal Asymmetric Price Transmission (APT) linear model among candidate ones. This study introduces variants of the Minimum Description Length criterion to APT modelling framework through a 1000 Monte Carlo simulations using a sample of 150 and compares them to the performances of widely used AIC and BIC. Results indicate that the performance of all criteria deteriorates with increasing noise across both standard and complex APT models. At higher noise level (3), eMDL and gMDL are alternatives to AIC which recovered strongest while nMDL is superior to all other criteria in recovering the true standard and complex models, respectively. The eMDL and gMDL outperformed all criteria (standard models) whilst nMDL and rMDL also outperformed all criteria (complex models) when the noise level was moderate (2). Lastly, at a lower noise level (1), rMDL was comparable to BIC which recovered strongest whiles nMDL is an alternative to AIC which was superior for the standard and complex models, respectively. Lower noise improves the performance of model selection methods in their ability to recover the true data generating process.