Portfolio Selection via Subset Resampling

Weiwei Shen, Jun WangAAAI 2017 · DOI

Abstract

Standard mean-variance portfolio optimization often fails in practical applications because it tends to maximize estimation errors in asset returns and risks. This paper introduces Subset Resampling (SSR), an ensemble method that constructs a diversified global portfolio by averaging numerous sub-portfolios optimized on small, accurately estimated asset subsets. SSR addresses the dimensionality problem inherent in high-dimensional financial data, providing a robust alternative to traditional benchmarks and modern machine learning strategies. Empirical results across four large-scale datasets demonstrate that SSR consistently achieves superior risk-adjusted returns with significantly lower trading costs and volatility.

Introduction

Since its introduction in 1952, Markowitz’s mean-variance optimization has been the cornerstone of modern portfolio theory. The model aims to find asset weights that minimize risk for a target return. However, in practice, this model is notoriously fragile due to estimation risk. To implement the model, one must estimate mean returns and the covariance matrix from historical data. When the number of assets ($n$) is large relative to available data, these estimates become noisy. The optimization process then "maximizes" these errors, leading to extreme, unstable asset weights that perform poorly out-of-sample. While previous attempts like Michaud’s "Resampled Efficient Portfolio" used bootstrapping to mitigate this, they often failed to address the underlying dimensionality bottleneck.

Methods

Subset Resampling (SSR) draws inspiration from the Random Subspace Method used in machine learning (e.g., Random Forests). Instead of attempting to optimize the entire universe of assets at once, SSR follows a four-step ensemble pipeline:

  1. Subsetting: Randomly select a small subset of $b$ assets from the total universe of $n$ assets.
  2. Local Optimization: Estimate the mean and covariance for just those $b$ assets and calculate their optimal weights.
  3. Repetition: Repeat this process $s$ times (typically thousands of iterations).
  4. Aggregation: Average the weights from all sub-portfolios to create the final, well-diversified global portfolio.

By reducing the number of assets in each sub-calculation, the estimation of the covariance matrix becomes significantly more stable and accurate. The final averaging step then restores the diversification benefits that were temporarily sacrificed at the subset level. Notably, traditional "Bagging" (resampling data points) is often ineffective when the number of assets is the bottleneck; resampling the asset subspaces is far more effective for high-dimensional financial data.

Figure 1: Efficient frontiers showing the impact of subset size (b) and number of resamples (s) on estimated vs. actual performance.

Results and Discussion

The authors evaluated SSR against ten competing strategies across four diverse datasets: FF100 (monthly returns), ETF139 (weekly returns), EQ181 (weekly returns), and SP434 (daily returns). SSR demonstrated marked superiority across almost all evaluation criteria.

Figure 3: Performance comparison of SSR against competing strategies across four major datasets. SSR (bolded) shows superior Sharpe Ratios and lower risk metrics.

Key Performance Insights

  • Sharpe Ratio (SR): SSR consistently achieved the highest risk-adjusted returns. For example, on the FF100 dataset, SSR reached a Sharpe Ratio of 1.51, compared to 0.93 for the standard Equally Weighted (EW) strategy and 0.62 for Michaud’s Resampled (RES) strategy.
  • Volatility and Drawdown: SSR exhibited significantly lower annualized volatility and Maximum Drawdown (MDD) than most competitors, indicating it is a safer strategy during market turbulence.
  • Turnover Rate (TO): A major practical advantage is SSR’s low turnover. Active strategies often fail due to high trading fees; SSR’s turnover was frequently an order of magnitude lower than other high-performing strategies.

The Precision-Diversification Tradeoff

There is a critical "sweet spot" for the subset size ($b$). If $b$ is too small, the portfolio loses diversification benefits; if $b$ is too large, estimation errors return. The authors suggest a rule of thumb: $b$ should be at least 30 for diversification and ensure at least 5 data points per asset for stable estimation. SSR acts as a natural regularizer; by averaging many sub-portfolios, the extreme weights that plague standard models are smoothed out into a balanced allocation.

Conclusion

Subset Resampling provides a robust framework for high-dimensional portfolio selection by leveraging ensemble stability to mitigate estimation risk. The method's superior performance and low turnover make it a highly practical tool for real-world investment. Future work may focus on integrating transaction costs and market frictions directly into the optimization objective to further refine the strategy for high-frequency environments.

Summary updated

BibTeX
@inproceedings{shen2017portfolio,
  author = {Weiwei Shen and Jun Wang},
  title = {{Portfolio Selection via Subset Resampling}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.1609/aaai.v31i1.10728},
}
Figure