Portfolio Blending via Thompson Sampling

Weiwei Shen, Jun WangIJCAI 2016

Abstract

Modern Portfolio Theory (MPT) often suffers from high sensitivity to estimation errors, leading to poor out-of-sample performance. This paper proposes a novel online learning framework that utilizes Thompson Sampling to dynamically blend basis portfolios. By framing the selection of blending weights as a multi-armed bandit problem, the proposed TS-EM and TS-VM strategies adapt to market conditions in real-time. Empirical evaluations across five diverse datasets demonstrate that this approach consistently outperforms traditional benchmarks, including the equally-weighted "1/n" strategy and Markowitz’s mean-variance model, in terms of risk-adjusted returns.

Introduction

The cornerstone of quantitative finance, Harry Markowitz’s Modern Portfolio Theory, relies heavily on the accurate estimation of future returns and risk. However, the "estimation error" inherent in these inputs often results in erratic portfolio weights and suboptimal performance in real-world applications. To address this, researchers have turned to portfolio blending—combining different strategies to enhance robustness.

The primary challenge in blending is determining the optimal weights (coefficients) dynamically. Static weights lack the flexibility required for volatile markets, while complex optimization can re-introduce estimation sensitivity. This paper introduces an online machine learning approach using Thompson Sampling to solve the blending weight "conundrum."

Methodology

The authors treat portfolio blending as a Bernoulli bandit problem, solved via Thompson Sampling (TS).

Basis Portfolios

The framework utilizes three standard financial building blocks:

  1. Equally-weighted (EW): The "1/n" strategy, assigning equal weight to all assets.
  2. Value-weighted (VW): Weights based on market capitalization, tracking the broader market.
  3. Minimum-variance (MV): A strategy focused exclusively on risk minimization.

The Thompson Sampling Mechanism

Two blended portfolios are proposed: TS-EM (blending EW and MV) and TS-VM (blending VW and MV). The blending coefficient ($\delta$) is modeled using a Beta distribution ($Beta(a, b)$). In each trading period, a weight is sampled from the distribution and its performance is compared against a benchmark weight (the distribution mean). If the sampled weight outperforms the benchmark, the parameters $a$ and $b$ are updated using Bayesian rules. This allows the algorithm to "learn" the market's preference for specific blends while balancing exploration and exploitation.

Experimental Setup

The proposed methods were evaluated against five datasets representing various market conditions and asset classes.

Table 1: Description of the five datasets used for evaluation, including Fama-French portfolios and real-world market data.

As shown in Table 1, the datasets include three standard Fama-French (FF) portfolios spanning over 40 years, and two recent market datasets (ETF139 and EQ181) covering the high-volatility period following the 2008 financial crisis.

Results and Discussion

The performance of TS-EM and TS-VM was compared against seven state-of-the-art strategies. Key metrics included the Sharpe Ratio (SR), Volatility (Vol), and Maximum Drawdown (MDD).

Table 2: Performance comparison across datasets. TS-EM and TS-VM consistently achieve superior Sharpe Ratios and lower risk metrics compared to benchmarks.

Key Findings

  • Superior Risk-Adjusted Returns: The Thompson Sampling-based portfolios achieved the highest Sharpe Ratios in nearly all datasets. For instance, on the FF100 dataset, TS-VM reached a Sharpe Ratio of 37.90%, outperforming the EW (26.75%) and VW (29.76%) benchmarks.
  • Robustness to Volatility: The blended portfolios generally exhibited lower volatility and smaller maximum drawdowns than their individual components, demonstrating that adaptive blending effectively mitigates the weaknesses of standalone strategies.
  • Statistical Significance: Bootstrapping analysis confirmed that the performance gains were statistically significant and not merely the result of market noise.

Conclusion

This research demonstrates that Thompson Sampling provides a powerful, adaptive framework for portfolio blending. By treating weight selection as a learning problem, the proposed strategies overcome the limitations of static blending and traditional optimization. While currently limited to two-fund blends, future work involving Dirichlet distributions could extend this framework to multi-portfolio blending, further enhancing its utility in complex financial environments.

Summary updated

BibTeX
@inproceedings{shen2016portfolio,
  author = {Weiwei Shen and Jun Wang},
  title = {{Portfolio Blending via Thompson Sampling}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure