Forecasting time-varying daily betas: a new nonlinear approach

Abstract

Purpose – The purpose of this paper is to examine the predictive ability of different well-known models for capturing time variation in betas against a novel approach where the beta coefficient is treated as a function of market return. Design/methodology/approach – Different GARCH models, the Kalman filter algorithm and the Schwert and Seguin model are used against our novel approach. The mean square error, the mean absolute error and the Diebold and Mariano test statistic constitute the measures of forecast accuracy. All models are tested over nine consecutive years and three different samples. Findings – The results show substantial differences in predictive accuracy among the samples. The new approach of modelling the systematic risk overwhelms the rest of the models in longer samples. In the smallest sample, the Kalman filter random walk model prevails. The examination of parameters between two groups of stocks with best and worst accuracy results depicts significant variations. For these stocks, the iid assumption of return is rejected and large differences exist on diagnostic tests. Originality/value – This study contributes to the literature with different ways. First, it examines the predictive accuracy of betas with different well-known models and introduces a novel approach. Second, after constructing betas from the estimated models’ parameters, they are used for out-of-sample instead of in-sample forecasts over nine consecutive years and three different samples. Third, a more closely examination of the models’ parameters could signal at an early stage the candidate models with the expected lowest forecasting errors. Finally, the study carries out some diagnostic tests for examining whether the existence of iid normal returns is accompanied by better performance.