Portfolio Construction Using Predictive Linear Model – An Adaptive Multi-Objective Approach

Abstract

The purpose of this study is to incorporate some of the influential findings in the forecasting literature in an integrated framework to examine whether a real-time optimizing investor can benefit from the stock market by allocating assets based on a predictive model that only uses industry portfolio and interest rate data as predictors. The proposed method aims to allow economic performance measures to have an impact on all steps of model building from variable selection to model combination without undermining the statistical performance measure. The predictors/models are selected from almost 300 variables by a multi-objective genetic algorithm considering both statistical and economic measures. I chose a subset of models from the Pareto-optimal frontier using a number of heuristic methods from the multi-criteria decision making (MCDM) literature and the concept of knee-point of a curve. The forecast of the next period is obtained by combining the forecasts of the selected subset of predictive linear regressions using a Bayesian variable selection and model averaging method called Stochastic Search Variable Selection (SSVS). The investor’s utility function is used to obtain the weight of a risky asset based on the output of the forecasting model. All aforementioned steps only use data up to the current time and before the forecasting time. The results are compared to common benchmarks such as the buy-and-hold strategy and additional benchmarks that are based on the findings of previous literature. The findings indicate that using the proposed approach can improve the portfolio performance measures relative to all benchmarks.

Keywords: Forecasting stock market, Adaptive forecasting, Dynamic variable selection, Multi-objective genetic algorithm, Economic and statistical performance measure

JEL Classification: C11; C53; G11