Multiple Portfolio Blending Strategy with Thompson Sampling

Mean-variance portfolio is one of the most practically used portfolios for the purpose of portfolio construction and asset allocation. Due to some problems such as inaccurate estimation of parameters, however, it actually ends up ineffective in terms of mean-variance, which is demonstrated by the widely accepted fact that heuristic strategies such as risk-based portfolio and equal weight portfolio perform better than it. In general, the most efficient one among these heuristic portfolios is difficult to find out, while the method to achieve more stable performance by blending multiple portfolios effectively is yet to be discussed well enough. In this paper, we present a practical online algorithm to blend multiple heuristic portfolios in the appropriate ratio by regarding it as multi-armed bandit problem. Specifically, we expand the algorithm proposed by a previous study, which leverages Thompson sampling to determine the blending ratios of two portfolios, so that our new algorithm can be applied to blend three or more portfolios to good effect. Here we assume the performance of each portfolio follows a Dirichlet distribution, and in order to keep each distribution online confine the previous data used to generate them within most recent return of each portfolio. The empirical analysis through the heuristic portfolios constructed with some representative benchmarks illustrates the superiority of our method over conventional trading strategies.