On Universal Portfolios with Continuous Side Information

Alankrita Bhatt, J. Jon Ryu, Young-Han Kim

Abstract

A new portfolio selection strategy that adapts to a continuous side-information sequence is presented, with a universal wealth guarantee against a class of state-constant rebalanced portfolios with respect to a state function that maps each side-information symbol to a finite set of states. In particular, given that a state function belongs to a collection of functions of finite Natarajan dimension, the proposed strategy is shown to achieve, asymptotically to first order in the exponent, the same wealth as the best state-constant rebalanced portfolio with respect to the best state function, chosen in hindsight from observed market. This result can be viewed as an extension of the seminal work of Cover and Ordentlich (1996) that assumes a single-state function.

Introduction

The concept of Universal Portfolios, pioneered by Thomas Cover in the 1990s, revolutionized the approach to sequential investment. By guaranteeing performance nearly as good as the best Constant Rebalanced Portfolio (CRP) in hindsight—without requiring prior knowledge of market distributions—Universal Portfolios provided a robust framework for online learning in financial markets. Subsequent extensions introduced "side information," allowing strategies to adapt based on external signals. However, these extensions typically assumed discrete side information (e.g., binary volatility states).

In real-world finance, signals such as interest rates, price-to-earnings ratios, and historical price relatives are inherently continuous. This paper addresses the fundamental challenge of adapting universal portfolios to continuous-valued side information, bridging the gap between portfolio optimization and modern statistical learning theory.

The Complexity of Continuous Signals

Mapping continuous side information into actionable trading states presents a significant theoretical hurdle. A mapping that is too coarse loses valuable predictive information, while a mapping that is too granular risks overfitting to noise. Previous models relied on a finite set of states, but continuous data offers an infinite variety of potential signals.

To bridge this gap, the authors leverage the Natarajan dimension, a multi-class generalization of the Vapnik-Chervonenkis (VC) dimension. By considering a class of functions $\mathcal{G}$ that map continuous inputs to discrete states, the researchers can bound the complexity of the side-information processing. This ensures that the portfolio strategy remains tractable and avoids the pitfalls of infinite-dimensional learning.

Methodology: Epochs and Empirical Coverings

The proposed strategy employs an innovative algorithmic structure based on several technical pillars:

  1. Log-Loss Equivalence: The strategy exploits the mathematical duality between maximizing cumulative wealth and minimizing log-loss in sequential probability estimation. This allows the portfolio problem to be framed as an online learning task.
  2. Epoch-Based Learning: Rather than attempting to learn from the entire data stream at once, the algorithm divides time into exponentially increasing "epochs." This structure allows the model to refine its understanding of the side information as more data becomes available.
  3. Empirical Covering: At the start of each epoch, the algorithm constructs an "empirical covering"—a representative subset of functions from the class $\mathcal{G}$. This subset effectively summarizes the possible ways to interpret the continuous data observed so far.
  4. Mixture Strategy: The final portfolio is constructed as a "mixture" or a fund-of-funds. It aggregates the performance of multiple universal portfolios, each operating under a different state-mapping function derived from the empirical covering.

Theoretical Insights

A key finding of this research is the necessity of stochastic assumptions when dealing with continuous side information. While discrete side information can be managed in a purely adversarial (worst-case) setting, the authors demonstrate that continuous signals require mild stochasticity, such as stationarity or $\beta$-mixing (where the process's memory fades over time). Without these assumptions, the "Littlestone dimension" of the function class becomes infinite, making universal learning theoretically impossible.

The authors prove that their strategy achieves theoretical universality: the expected average regret—the gap between the strategy's wealth and the best possible state-constant rebalanced portfolio chosen in hindsight—vanishes as the number of trading days $n$ approaches infinity.

Empirical Results

The strategy was validated using both synthetic stochastic processes and real-world market data. In a case study involving two stocks, Ford (F) and Macy’s (M), over a six-year period (2012–2017, 1508 trading days), the proposed strategy used the previous day’s price relatives as continuous side information.

The results demonstrated clear outperformance:

  • Ford (Buy & Hold): 1.11x growth
  • Macy’s (Buy & Hold): 0.77x growth
  • Standard Universal Portfolio (No side info): 0.98x growth
  • Proposed Strategy: 1.15x growth

By effectively "quantizing" the continuous price signals into actionable states, the proposed strategy outperformed both the individual assets and the traditional universal portfolio.

Limitations and Future Work

While the theoretical results are robust, practical implementation faces challenges. The computational cost of finding a minimal empirical covering for high-dimensional function classes is significant. In the experiments, the authors utilized simple 1D threshold functions; scaling this to complex, multi-dimensional economic indicators remains a bottleneck.

Furthermore, the current model assumes a frictionless market. In practice, the frequent rebalancing required by a mixture strategy would incur transaction costs that could diminish net returns. Finally, extending the framework to handle adversarial side information—without the requirement of stationarity—remains an open and challenging area for future research.

Summary updated

Abstract

A new portfolio selection strategy that adapts to a continuous side-information sequence is presented, with a universal wealth guarantee against a class of state-constant rebalanced portfolios with respect to a state function that maps each side-information symbol to a finite set of states. In particular, given that a state function belongs to a collection of functions of finite Natarajan dimension, the proposed strategy is shown to achieve, asymptotically to first order in the exponent, the same wealth as the best state-constant rebalanced portfolio with respect to the best state function, chosen in hindsight from observed market. This result can be viewed as an extension of the seminal work of Cover and Ordentlich (1996) that assumes a single-state function.
BibTeX
@inproceedings{bhatt2023universal,
  author = {Alankrita Bhatt and J. Jon Ryu and Young-Han Kim},
  title = {{On Universal Portfolios with Continuous Side Information}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2023},
  url = {https://openreview.net/forum?id=XWjGGFhnt0},
}
Figure