Optimal Mean-Reverting Portfolio Design: Balancing Convergence and Profitability

Executive Summary

In the world of Statistical Arbitrage (Stat Arb), the "holy grail" is constructing a synthetic asset that is both highly stationary and sufficiently volatile. This paper, published by Zhao et al., introduces a robust optimization framework for the Mean-Reverting Portfolio (MRP) design. By moving beyond the limitations of L2-norm and simple budget constraints, the authors provide a mathematically rigorous way to handle leverage constraints (L1-norm) while maximizing trading opportunities and profit potential.

The Core Conflict: Mean Reversion vs. Variance

Quantitative traders face a fundamental trade-off:

1. Mean Reversion (MR): Higher MR means the spread crosses its equilibrium more frequently, providing more "entries" and "exits."
2. Variance: Higher variance suggests that when the price deviates from the mean, the potential profit per trade is larger.

Previous SOTA methods often maximized one at the expense of the other or utilized constraints that didn't translate well to real-world margin requirements. The authors' insight is that by formulating this as a multi-objective problem under a strict leverage cap, they can find a Pareto-optimal frontier of spreads.

Methodology: The SCA-MRP Framework

The objective function is notoriously difficult because MR criteria (like Predictability or Portmanteau statistics) are typically ratios of quadratic forms, making the problem non-convex.

1. Successive Convex Approximation (SCA)

Instead of attacking the non-convex problem directly, the authors use SCA. At each iteration $k$, they approximate the complex objective with a surrogate function that is:

• Strongly Convex: Ensuring a unique global minimum for each sub-step.
• Tangent to the Original: Preserving the gradient information at the current iterate.

2. The Leverage Constraint

The paper defines the leverage set as: $\mathcal{W}riangleq\{\mathbf{w}\mid \Vert \mathbf{Bw}{\Vert }_1\le L\}$ where $L$ is the total dollar position allowed. This is a critical departure from earlier works as it directly maps to an investor's capital allocation limits.

Figure 1: The full pipeline from asset selection and cointegration analysis to MRP design and final trading execution.

3. Efficient Subproblem Solving

To handle the inner loop of the SCA, the authors developed:

• ADMM: For general cases.
• Majorized ADMM (M-ADMM): To achieve closed-form updates for the weights.
• Water-Filling Projections: An $O(N\log N)$ algorithm to project weights onto the L1-ball.

Experimental Validation

The authors tested their framework using a Vector Error Correction Model (VECM) on both synthetic data and a basket of S&P 500 stocks (e.g., IBM, MMM, AXP).

Figure 2: Performance comparison of the proposed SCA-MRP variations against fmincon. Notice the significantly faster convergence and lower final objective value.

Key Findings:

• Efficiency: The specialized SCA-MRP (using M-ADMM) outperformed general solvers like MOSEK and fmincon by orders of magnitude in terms of runtime.
• Practicality: Unlike the "MRP with budget" used in prior literature, the "MRP with leverage" formulation allows traders to tune the parameter $\mu$ to find a spread that matches their risk tolerance without violating margin limits.
• ROI and Sharpe Ratio: Results on real market data (2013-2014) showed that the optimized MRP consistently outperformed individual cointegrated spreads (Spread $s_1$) and legacy L2-norm models.

Figure 3: Cumulative P&L and Sharpe Ratio for the proposed MRP versus a standard single spread ($s_1$).

Critical Insight & Conclusion

The genius of this work lies in its Heuristic-to-Physical translation. By reformulating the mathematical "regularization" into a "leverage constraint," the authors made the optimization problem relevant to the floor traders of Wall Street.

Takeaway: Effective Stat Arb is not just about finding stationarity; it’s about finding governed stationarity. For developers and researchers, the SCA-MRP library offers a template for converting complex non-convex signal processing problems into high-performance, real-time trading systems.

Limitations: While the framework is robust, it assumes a constant cointegration space. In highly volatile markets, the "cointegration space" itself may shift (regime switching), necessitating an adaptive or online version of this algorithm.