PAMR: Passive aggressive mean reversion strategy for portfolio selection

Bin Li, Peilin Zhao, Steven C. H. Hoi, Vivekanand GopalkrishnanMachine Learning 2012 · OpenReview

Abstract

This article proposes a novel online portfolio selection strategy named “Passive Aggressive Mean Reversion” (PAMR). Unlike traditional trend following approaches, the proposed approach relies upon the mean reversion relation of financial markets. Equipped with online passive aggressive learning technique from machine learning, the proposed portfolio selection strategy can effectively exploit the mean reversion property of markets. By analyzing PAMR’s update scheme, we find that it nicely trades off between portfolio return and volatility risk and reflects the mean reversion trading principle. We also present several variants of PAMR algorithm, including a mixture algorithm which mixes PAMR and other strategies. We conduct extensive numerical experiments to evaluate the empirical performance of the proposed algorithms on various real datasets. The encouraging results show that in most cases the proposed PAMR strategy outperforms all benchmarks and almost all state-of-the-art portfolio selection strategies under various performance metrics. In addition to its superior performance, the proposed PAMR runs extremely fast and thus is very suitable for real-life online trading applications.

Introduction

In quantitative finance, portfolio selection involves the dynamic distribution of wealth across multiple assets to maximize cumulative returns. Existing strategies generally fall into two categories: trend following, which bets on the persistence of price movements, and regret minimization, which seeks to perform as well as the best fixed portfolio in hindsight. However, trend following often fails in range-bound markets, and many regret-minimization algorithms suffer from high computational complexity.

The Passive Aggressive Mean Reversion (PAMR) strategy addresses these challenges by leveraging the mean reversion principle—the financial observation that asset prices tend to return to their long-term averages. By integrating this principle with Passive Aggressive (PA) machine learning techniques, PAMR offers a computationally efficient, online approach to portfolio management.

Methodology

The core of PAMR is the adaptation of the Passive Aggressive learning framework to the domain of portfolio selection. The algorithm operates on a daily update cycle, adjusting the portfolio based on the previous day's performance.

The Mean Reversion Principle

The fundamental intuition behind PAMR is "buying low and selling high." When an asset's price drops significantly relative to the market, it is expected to bounce back; conversely, an outperforming asset is expected to revert.

Table 1: Illustration of BCRP exploiting mean reversion through rebalancing.

As shown in Table 1, a Best Constant Rebalanced Portfolio (BCRP) exploits mean-reverting price sequences by systematically moving wealth from "winners" to "losers." PAMR formalizes this intuition through a rigorous optimization framework.

The PAMR Algorithm

PAMR utilizes an $\epsilon$-insensitive loss function. If the portfolio's return is below a threshold $\epsilon$, the algorithm remains "passive." If the return exceeds $\epsilon$, indicating a potential peak in the current "winners," the algorithm becomes "aggressive" and rebalances.

A critical feature of the PAMR update rule is the risk-adaptive step size ($\tau_t$). The denominator of this step size is proportional to the market's variance. In highly volatile or "noisy" environments, the step size decreases, preventing the algorithm from overreacting to random price fluctuations. To further enhance robustness, variants like PAMR-1 and PAMR-2 introduce slack variables to handle extreme price spikes without destabilizing the portfolio.

Experimental Evaluation

The authors evaluated PAMR against 12 competing strategies across six diverse real-world datasets, including the NYSE (Old and New), Toronto Stock Exchange (TSE), S&P 500, MSCI global indices, and the Dow Jones Industrial Average (DJIA).

Performance and Efficiency

  • Superior Returns: On the NYSE (O) dataset, PAMR achieved a theoretical cumulative wealth of $5.14 \times 10^{15}$ (starting from $1), significantly outperforming the market index ($14.5) and the best-in-hindsight BCRP ($250.6).
  • Robustness: PAMR remained profitable even when accounting for transaction costs up to 0.5%, demonstrating its practical viability.
  • Computational Speed: PAMR exhibits linear time complexity $O(mn)$, where $m$ is the number of stocks and $n$ is the number of trading days. In backtests, it processed datasets in under 8 seconds, whereas some competitors required thousands of seconds.

Conclusion

PAMR represents a significant advancement in online portfolio selection, combining the financial logic of mean reversion with the algorithmic efficiency of Passive Aggressive learning. While its performance is dependent on the presence of mean-reverting behavior—as seen in its lower performance on the DJIA dataset—its speed and robustness make it a compelling tool for real-time trading applications. Future research may focus on hybrid models that can dynamically switch between mean reversion and trend-following strategies based on market conditions.

Summary updated

Abstract

This article proposes a novel online portfolio selection strategy named “Passive Aggressive Mean Reversion” (PAMR). Unlike traditional trend following approaches, the proposed approach relies upon the mean reversion relation of financial markets. Equipped with online passive aggressive learning technique from machine learning, the proposed portfolio selection strategy can effectively exploit the mean reversion property of markets. By analyzing PAMR’s update scheme, we find that it nicely trades off between portfolio return and volatility risk and reflects the mean reversion trading principle. We also present several variants of PAMR algorithm, including a mixture algorithm which mixes PAMR and other strategies. We conduct extensive numerical experiments to evaluate the empirical performance of the proposed algorithms on various real datasets. The encouraging results show that in most cases the proposed PAMR strategy outperforms all benchmarks and almost all state-of-the-art portfolio selection strategies under various performance metrics. In addition to its superior performance, the proposed PAMR runs extremely fast and thus is very suitable for real-life online trading applications. The experimental testbed including source codes and data sets is available at http://www.cais.ntu.edu.sg/~chhoi/PAMR/ .
BibTeX
@article{li2012pamr,
  author = {Bin Li and Peilin Zhao and Steven C. H. Hoi and Vivekanand Gopalkrishnan},
  title = {{PAMR: Passive aggressive mean reversion strategy for portfolio selection}},
  journal = {Machine Learning},
  year = {2012},
  url = {https://openreview.net/forum?id=SfL9XwUOIGI},
}

Loading related content...

Figure