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
In the realm of quantitative finance, the quest for optimal portfolio selection often grapples with the unpredictability of market distributions. Traditional log-optimal portfolios rely on the assumption that markets follow known, stationary probability distributions—an assumption frequently violated in practice. To address this, Thomas Cover introduced the concept of Universal Portfolios in 1991. These strategies do not assume a specific market model; instead, they provide a mathematical guarantee that the investor's wealth will asymptotically match the exponential growth rate of the best Constant Rebalanced Portfolio (CRP) chosen in hindsight.
Subsequent research by Cover and Ordentlich (1996) expanded this framework to include "side information," such as discrete signals (e.g., "High Volatility" vs. "Low Volatility"). However, real-world side information—such as historical price relatives or macroeconomic indicators—is typically continuous. Until now, incorporating such continuous signals required manual discretization or restrictive modeling assumptions. This paper, authored by researchers from UC San Diego and Gauss Labs, introduces a rigorous framework for universal portfolios that adapt to continuous side-information sequences.
Methodology
The proposed strategy bridges information theory and financial mathematics through three primary innovations:
1. State Functions and Mapping
The authors define a class of state functions, $\mathcal{G}$, which map continuous side-information symbols (e.g., a vector of the previous day's prices) into a finite set of discrete states. This allows the portfolio to switch between different CRPs based on the current state of the continuous signal.
2. Complexity Control via Natarajan Dimension
A significant challenge in using complex side information is the risk of overfitting. To manage this, the authors utilize the Natarajan dimension, a multi-class generalization of the Vapnik–Chervonenkis (VC) dimension. By requiring the class of state functions to have a finite Natarajan dimension, the researchers establish a mathematical "speed limit" on the complexity of the side information the algorithm can process without being overwhelmed by noise.
3. Epoch-Based Mixture Strategy
Rather than attempting to identify the optimal state function immediately, the algorithm employs an epoch-based approach. Time is divided into doubling epochs. During each epoch, the strategy evaluates the performance of all possible state functions from the previous epoch and constructs a "mixture" portfolio. This bootstrapping mechanism allows the strategy to learn the most effective way to partition continuous information into states over time.
Theoretical Insights
A pivotal contribution of this work is the formalization of the connection between portfolio optimization and information theory. In Proposition 1, the authors prove that portfolio selection is mathematically equivalent to a "log-loss" prediction problem. Effectively, if an algorithm can accurately predict market weights in terms of log-loss, it will naturally maximize cumulative wealth.
The authors demonstrate that as the number of trading days $n$ approaches infinity, the "regret"—the difference in log-wealth between the proposed algorithm and the best possible state-dependent strategy chosen in hindsight—diminishes to zero. This establishes the theoretical universality of the approach under the assumption that side information follows a stationary stochastic process.
Empirical Evaluation
While primarily a theoretical contribution, the paper includes an empirical test using historical stock data for Ford (F) and Macy’s (M) from 2012 to 2017 (1,508 trading days). The "continuous signal" provided to the algorithm was the price relative of the first stock from the preceding day.
The results highlighted the efficacy of the continuous side-information approach:
- Best CRP in Hindsight: A buy-and-hold strategy for Ford yielded a 1.11x growth factor.
- Standard Universal Portfolio: Without side information, the standard universal strategy underperformed a simple uniform split.
- Proposed Strategy: By leveraging continuous side information, the authors' strategy achieved a 1.15x growth factor, outperforming the best constant strategy by dynamically reacting to price movements.
Limitations and Future Directions
Despite its theoretical robustness, the strategy faces practical hurdles. The computational complexity of calculating these portfolios is $O(T^m)$, where $T$ is time and $m$ is the number of assets, necessitating Monte Carlo approximations for simulations. Furthermore, the current proof assumes a stationary process for side information; whether this universality holds in purely "adversarial" (non-stochastic) environments remains an open question. Finally, the model does not yet account for transaction costs or market impact, which are critical factors for real-world implementation.
Summary updated