Low Volatility Stock Portfolio Through High Dimensional Bayesian Cointegration

Parley Ruogu Yang, Alexander Y. ShestopaloffarXiv 2024 · DOI · arXiv

TL;DR

This paper introduces a high-dimensional Bayesian framework to identify "cointegration"—hidden stationary relationships—among hundreds of stocks simultaneously. By using a Spike-and-Slab Lasso (SSL) prior to enforce sparsity, the authors construct long-short portfolios that exhibit remarkably low volatility and significantly improved risk-adjusted returns compared to standard market benchmarks.

What Problem Does This Paper Address?

In finance, most stock prices are non-stationary, meaning their statistical properties like mean and variance change over time. While traders often look at price changes (returns), "cointegration" looks for specific combinations of stocks that, when held together, create a stationary (mean-reverting) portfolio.

The massive hurdle here is the "curse of dimensionality." If you want to find these relationships among the S&P 500, a standard Vector Autoregressive (VAR) model would require estimating roughly 250,000 parameters. To do this reliably with classical statistics, you would need nearly 1,000 years of daily data—obviously an impossible requirement for modern markets. Consequently, most cointegration research is limited to small "pairs trading" (2-3 stocks) rather than large-scale index optimization.

Methods

The authors bridge the gap between high-dimensional data and cointegration theory using a Bayesian Decomposed Vector Error Correction Model (VECM).

  1. Decomposition: They simplify the estimation of the impact matrix ($\Pi$) by decomposing it into an upper triangular matrix ($R$) and an orthogonal matrix ($S$). The rank of $R$ tells us how many independent cointegrated portfolios exist.
  2. Spike-and-Slab Lasso (SSL): To handle the 500+ features with only a few years of data, they apply an SSL prior. This acts as a sophisticated "filter": the "spike" pushes unimportant relationships toward zero, while the "slab" allows significant relationships to remain.
  3. EM Algorithm: They employ an Expectation-Maximization (EM) algorithm to iteratively solve for the sparsest, most stable version of the portfolio matrix.
  4. Portfolio Construction: The resulting non-zero rows of the estimated matrix become the weights for long-short portfolios. These are then "unified" (normalized) so the total absolute weight equals one.

Datasets

The study evaluates three distinct asset spaces using rolling training and testing windows (2021–2024):

  • S&P 500 (SPY): 453 constituent stocks identified as non-stationary ($I(1)$).
  • STOXX Europe 600 (SXXP): 544 constituent stocks.
  • Joint Universe: A combined basket of 997 US and European stocks.

The models were trained on approximately two years of daily data (~500 observations) and tested out-of-sample on the subsequent month.

Key Insights and Ideas

  • Sparsity is a Feature, Not a Bug: By forcing the model to find the simplest relationships (sparse rows), the authors ensure the portfolios are more likely to represent genuine economic ties rather than statistical noise.
  • Persistence of Stationarity: A common critique of cointegration is that it "breaks" quickly. This paper demonstrates that Bayesian-estimated relationships persist well into out-of-sample periods, maintaining their low-volatility properties even when the broader market is turbulent.
  • The "Hedge" Effect: The authors found that these cointegrated portfolios don't necessarily generate high returns on their own (they are often flat). However, when mixed with a passive index (like the S&P 500), they act as a powerful stabilizer, drastically reducing drawdowns without sacrificing the entire upside.

Key Results

The quantitative improvements over the benchmarks are substantial:

  • Volatility Reduction: While the US benchmark had a volatility of ~21%, the cointegrated portfolios often showed in-sample volatilities below 1%.
  • Sharpe Ratio Improvement: In out-of-sample testing, an optimized combination of the S&P 500 and cointegrated portfolios achieved a Sharpe ratio of 1.96, compared to 1.70 for the benchmark alone.
  • Drawdown Protection: The maximum drawdown for the optimized US portfolio was only 4.5%, whereas the benchmark suffered a 10.7% drop.
  • Extreme Stability: In a pure volatility-minimization exercise, the authors produced a joint US-European portfolio with a maximum drawdown of just 0.3%.

Limitations and Future Work

The authors acknowledge a few areas for improvement:

  • Partial Bayesianism: The current model relies on a partial least squares pre-estimation step. A "fully" Bayesian approach that estimates all parameters simultaneously could be more robust but is computationally more expensive.
  • Market Frictions: The study does not explicitly account for transaction costs, slippage, or currency hedging costs for the European stocks, which could eat into the performance of high-turnover long-short strategies.
  • Macroeconomic Shifts: While the relationships persist well, the model still faces challenges during major regime shifts (potential "change-points") in financial time series.

Future research may look into more complex Bayesian priors to better capture the "heavy tails" of financial data or integrate these models into more active high-frequency trading environments.

Summary updated

Abstract

We employ a Bayesian modelling technique for high dimensional cointegration estimation to construct low volatility portfolios from a large number of stocks. The proposed Bayesian framework effectively identifies sparse and important cointegration relationships amongst large baskets of stocks across various asset spaces, resulting in portfolios with reduced volatility. Such cointegration relationships persist well over the out-of-sample testing time, providing practical benefits in portfolio construction and optimization. Further studies on drawdown and volatility minimization also highlight the benefits of including cointegrated portfolios as risk management instruments.
BibTeX
@misc{yang2024low,
  author = {Parley Ruogu Yang and Alexander Y. Shestopaloff},
  title = {{Low Volatility Stock Portfolio Through High Dimensional Bayesian Cointegration}},
  year = {2024},
  eprint = {2407.10175},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.10175},
}
Figure