Abstract
This paper presents a novel Approximate Dynamic Programming (ADP) approach for multi-asset portfolio optimization under proportional transaction costs. By approximating the complex, high-dimensional "no-trading region" with a Löwner-John ellipsoid, the proposed method overcomes the curse of dimensionality that limits traditional approaches. Experimental results demonstrate that this geometric approximation achieves near-optimal risk-adjusted returns while providing a computational speedup of up to 500 times compared to standard benchmarks.
Introduction
In a frictionless market, the classic Merton portfolio strategy suggests maintaining a constant fraction of wealth in each asset. However, real-world transaction fees make frequent rebalancing prohibitively expensive. The optimal strategy in the presence of fees is to define a no-trading region: a zone where the portfolio is left alone, and trades are only executed when the allocation drifts outside this boundary.
Calculating the exact shape of this region for portfolios with many assets is a significant challenge. As the number of assets increases, the computational complexity grows exponentially—a phenomenon known as the "curse of dimensionality." Most existing solutions are restricted to one or two risky assets, making them impractical for modern, diversified portfolios.
Methods
The authors propose a hybrid approach that combines Value Function Iteration (VFI) with geometric policy parameterization. The process involves four key components:
- Efficient Sampling: Using Sobol low-discrepancy sequences to sample the state space more effectively than standard grids.
- Löwner-John Ellipsoid Approximation: The complex, multi-dimensional no-trading region is approximated by a Löwner-John ellipsoid—the unique minimum-volume ellipsoid that encloses the optimal post-trade positions calculated via VFI.
- Euclidean Projection: If a portfolio drifts outside the ellipsoid, the algorithm performs a simple Euclidean projection to move the allocation back to the nearest point on the ellipsoid's surface.
- Early Convergence: The researchers observed that these ellipsoids stabilize quickly after a few backward steps in time, allowing the algorithm to stop the expensive VFI process early and reuse the stable ellipsoid for earlier periods.
Results and Discussion
The method was evaluated using four distinct parameter settings, including synthetic benchmarks and real-world market data. The "Ellipsoid Approximation" (EA) method was compared against theoretical upper bounds and other strategies like Rolling Buy-and-Hold (RBH) and Myopic (MO) strategies.
Key findings include:
- Near-Optimality: The gap between the EA method and theoretical upper bounds was remarkably small, ranging from only 0.06% to 0.30% in annualized returns.
- Computational Efficiency: The EA method was 40 to 500 times faster than competing strategies, making it viable for high-dimensional portfolios.
- Superior Performance: In all tested datasets, the EA method consistently achieved the highest Annualized Certainty Equivalent (CER) among all lower-bound strategies.
Conclusion
By leveraging geometry as a shortcut, the authors demonstrate that a smooth ellipsoid approximation captures the essential features of optimal trading zones well enough to produce near-optimal results with massive speedups. While the current model focuses on proportional transaction costs, future work could extend this "VFI + geometric parameterization" framework to account for market impact, capital gains taxes, or non-convex trading regions.
Summary updated