Online Universal Dirichlet Factor Portfolios

Purushottam Parthasarathy, Avinash Bhardwaj, Manjesh K. HanawalarXiv 2023 · DOI · arXiv

Abstract

This paper enhances Thomas Cover’s classic "Universal Portfolio" strategy by replacing uniform random sampling with factor-informed Dirichlet sampling. By biasing portfolio weights toward known market drivers like "Size" or "Momentum," the researchers analytically and empirically prove they can generate significantly higher wealth and better risk-adjusted returns while maintaining the "universal" guarantee of performing as well as the best fixed-weight portfolio in hindsight.

Introduction: The Challenge of "Blind" Portfolios

In the world of online portfolio selection, the gold standard has long been Thomas Cover’s Universal Portfolio. The beauty of Cover’s approach is that it requires no statistical assumptions about the future; it simply averages all possible constant rebalanced portfolios (CRPs). Over time, this "blind" strategy is guaranteed to perform as well as the best fixed-weight portfolio you could have chosen if you had a crystal ball.

However, being "blind" is also a weakness. In the real world, we know that certain characteristics—factors—drive stock returns (e.g., small-cap stocks often outperform large-caps over long horizons). The challenge addressed in this research is: how do we inject this market intuition into a Universal Portfolio without losing the mathematical safety net of the "universal" property?

Methodology: The Factor Dirichlet Approach

The authors propose a Factor Dirichlet approach to bridge the gap between agnostic universal strategies and factor-based investing.

  1. Standard Universal Portfolio: Traditionally, a Universal Portfolio is built by sampling weights from a Uniform Dirichlet distribution ($\alpha = 1$). This treats every possible portfolio combination as equally likely at the start.
  2. Factor-weighted Dirichlet: The authors use a Dirichlet distribution where the $\alpha$ parameter is tied to specific asset factors. For example, the "Size" factor might use the inverse of the stock price to bias the distribution.
  3. Mathematical Foundation: The researchers establish a proof showing that if market returns follow a factor model, the wealth generated by these factor-tilted portfolios will "dominate" (strictly exceed) the wealth of the uniform version. Using the Cauchy-Schwartz inequality, they prove that the ratio of informed wealth to uninformed wealth is $\ge 1$.

Crucially, the Dirichlet distribution's $\alpha$ parameter acts as a "concentration" tool. By setting $\alpha$ based on factors, the algorithm is biased to spend more time exploring portfolio weights that align with historically successful economic theories, while still maintaining the "universal" property.

Experimental Setup

The approach was evaluated using two major equity indices:

  • Nasdaq 100: 71 assets that survived from August 2007 to June 2023.
  • FTSE 100: 87 assets from the London Stock Exchange with similar historical availability.

The researchers used weekly median prices to reduce noise and filtered for assets with at least 4,000 trading days of history.

Results: Performance and Risk-Adjusted Returns

The empirical results demonstrate that informed priors significantly enhance performance across different market environments.

NASDAQ 100 Performance

In the Nasdaq 100 backtest, the Momentum and Size-based Dirichlet portfolios significantly outperformed the standard Cover strategy in terms of total wealth growth. Risk-adjusted performance, measured by the Cumulative Moving Average Sharpe Ratio (CMAS), showed that almost all factor-tilted portfolios dominated the standard Cover portfolio.

Performance of Dirichlet Factor Portfolios on NASDAQ Constituents (2007-2023).

FTSE 100 Performance

In the FTSE 100 dataset, the "Size" factor (weighting toward lower-priced stocks) showed particularly outsized performance compared to all other allocations. This highlights the flexibility of the method; it is not tied to a single factor but can adapt to different market dynamics.

Performance of Dirichlet Factor Portfolios on FTSE Constituents (2007-2023).

Discussion and Future Work

While the results are promising, the authors acknowledge certain limitations:

  • Survivorship Bias: The study only included stocks that survived the entire 2007–2023 period. Real-world performance might be lower if delisted companies were included.
  • Factor Scope: The study focused on price-derived factors. Future work could incorporate fundamental accounting data (e.g., P/E ratios, book value) into the Dirichlet mixture.
  • Dynamic Learning: Currently, factor weights are somewhat static. A promising next step is "online learning" of the Dirichlet parameters, allowing the algorithm to learn which factors are most effective in real-time.

In conclusion, the Factor Dirichlet approach provides a robust framework for incorporating market intuition into universal portfolio strategies, offering a path toward higher returns without sacrificing theoretical guarantees.

Summary updated

BibTeX
@misc{parthasarathy2023online,
  author = {Purushottam Parthasarathy and Avinash Bhardwaj and Manjesh K. Hanawal},
  title = {{Online Universal Dirichlet Factor Portfolios}},
  year = {2023},
  eprint = {2308.07763},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.07763},
}
Figure