Portfolio Allocation with Time Series and Machine Learning Models

Abstract

The historical average stock return is commonly used to predict expected stock return in portfolio management for its better prediction and more straightforward calculations. However, most regressive models empirically proved to underperform the historical average return are explanatory models using predictor variables. This article takes the challenge to further explore the regressive power in predicting stock return but focuses on time series (TS) forecasting models and machine learning (ML) models. The results of this article show that many of these TS and ML regression models beat the historical average return, delivering higher portfolio Sharpe ratios and realized returns in backtesting. Although TS and ML models have inherently weaker explanatory power for their predictions, these results are still meaningful for retail and institutional mean-variance investors, opening up a new angle to forecast expected returns in portfolio allocation.