Asset selection via correlation blockmodel clustering

Wenpin Tang, Xiao Xu, Xun Yu ZhouExpert Systems with Applications 2022 · OpenReview

TL;DR

This paper introduces a novel data-driven clustering method, ACC, to identify a smaller, well-diversified set of stocks from a large universe. By formalizing diversification through a "correlation blockmodel," ACC offers a theoretically grounded and empirically superior approach to traditional clustering methods for portfolio construction.

What Problem Does This Paper Address?

Modern portfolio theory, pioneered by Markowitz, emphasizes diversification and the risk-return tradeoff. However, naively including all available assets in a portfolio is impractical due to costs and management complexity. Seasoned investors and large funds alike select a subset of assets. While traditional methods like industry classification exist, they can be subjective and may miss nuanced relationships. The core challenge this paper addresses is how to systematically select a smaller, yet sufficiently diversified, subset of stocks that achieves the benefits of a larger portfolio, moving beyond simple correlation-based clustering.

Methods

The paper proposes the Asset Clustering through Correlation (ACC) algorithm, built upon a novel correlation blockmodel. This model defines clusters based on two key criteria:

  1. High intra-cluster correlations: Assets within a cluster are highly correlated.
  2. Similar inter-cluster correlations: Assets within a cluster have similar correlation patterns with all other assets outside the cluster.

This second criterion is crucial and is formalized by the correlation blockmodel, which posits that asset returns within a group can be represented as a common latent factor plus uncorrelated noise. This ensures that assets within a cluster are interchangeable in terms of their relationship with external assets, a key aspect of diversification.

The ACC algorithm involves:

  • Estimating heavy-tailedness parameters: Financial data often exhibits heavy tails, so parameters (alpha and L) are estimated to account for this.
  • Defining a dissimilarity measure (CORD): This measures the difference in correlations of two assets with all other assets.
  • The PARTITION procedure: An iterative algorithm that groups assets based on the CORD measure and a threshold parameter (epsilon).
  • Threshold tuning: A data-driven approach using three rules to select the optimal epsilon, balancing intra-cluster correlation (Criterion 1) and cluster separation (Criterion 2).

Datasets

The study uses daily closing prices and historical constituent data for S&P 500 stocks from January 1996 to January 2020. The backtesting period for clustering and portfolio construction spans from February 2001 to January 2020.

Key Insights and Ideas

  • Diversification beyond industry: The ACC algorithm demonstrates that clustering based on correlation patterns can uncover meaningful groupings of stocks that transcend traditional industry classifications, identifying stocks with similar market movements regardless of their stated sector.
  • Interchangeability as a diversification metric: The core idea that assets within a cluster should have similar correlation patterns with all other assets is a powerful way to ensure diversification. This makes assets within a cluster effectively interchangeable for portfolio construction purposes.
  • Theoretical foundation for clustering: Unlike many heuristic clustering methods, ACC is grounded in a formal statistical model (the correlation blockmodel), providing a theoretical basis for its effectiveness.
  • Robustness to heavy tails: The explicit consideration and estimation of heavy-tailedness parameters in financial data makes the ACC algorithm more robust and reliable for real-world financial applications.

Key Results

  • Outperformance: Portfolios constructed using the ACC algorithm consistently outperform the S&P 500 ETF (SPY) and portfolios based on GICS sector/industry classifications across various performance metrics (Sharpe ratio, annualized return, volatility, drawdown).
  • Comparison with benchmarks: ACC outperforms k-medoids clustering in most scenarios, especially for risk parity and mean-variance allocation strategies. While k-medoids shows higher similarity to GICS sectors, ACC's distinctiveness is argued to be advantageous for uncovering "under-the-radar" stocks.
  • Long-term suitability: ACC portfolios are particularly well-suited for investors with long investment horizons, showing consistent outperformance over extended periods.
  • Efficiency with infrequent rebalancing: ACC portfolios perform well even with annual rebalancing, suggesting they are suitable for investors who prefer a "buy and hold" strategy with minimal trading.
  • Market-neutral portfolios: ACC-based market-neutral portfolios demonstrate resilience during market downturns, achieving positive returns and lower drawdowns than SPY.

Limitations and Future Work

  • Minimal separation: The paper acknowledges the challenge of recovering clusters when the minimal separation between clusters (min CORD) is below a certain threshold related to sample size and data properties. Future work could explore improving cluster recovery in such scenarios.
  • Integrating Criteria 1 & 2 into optimization: The authors suggest future research could directly embed Criteria 1 and 2 into mean-variance optimization algorithms, rather than using clustering as a pre-processing step.
  • Further algorithm improvements: The paper mentions the possibility of further refining the ACC algorithm.

Summary updated

Abstract

Highlights•Cluster financial assets to identify a small set of stocks of large diversification.•Develop a data-driven approach to clustering based on a correlation blockmodel.•Devise algorithms to detect clusters with theory and practical guidance.•Conduct empirical analysis to verify the performance of the algorithms.
BibTeX
@article{tang2022asset,
  author = {Wenpin Tang and Xiao Xu and Xun Yu Zhou},
  title = {{Asset selection via correlation blockmodel clustering}},
  journal = {Expert Systems with Applications},
  year = {2022},
  url = {https://openreview.net/forum?id=PViD8KENuX},
}
Figure