TL;DR
This paper introduces a new framework called "Linearity Generating Strategies" (LGS) that allows for dynamic portfolio optimization in the presence of realistic transaction costs and predictable asset returns. The key innovation is a closed-form solution that significantly simplifies a previously intractable problem, making it computationally efficient even with a large number of assets and complex return dynamics.
What Problem Does This Paper Address?
The classic Markowitz portfolio optimization framework, while foundational, often struggles to incorporate real-world complexities like transaction costs and time-varying expected returns. Academic research has largely ignored these frictions because they make the dynamic optimization problem computationally intractable, especially with many assets. Existing practical approaches often rely on ad-hoc adjustments. This paper bridges this gap by developing a tractable method for dynamic asset allocation that explicitly accounts for these realistic frictions, aiming to provide a more practical and performant solution.
Methods
The core contribution is the development of "Linearity Generating Strategies" (LGS). This approach restricts the set of possible trading strategies to a specific parametric class. Within this class, the complex, non-linear dynamic optimization problem is transformed into a deterministic linear-quadratic optimization problem, which can be solved efficiently in closed-form. The LGS framework allows for:
- General factor structures in returns: Accommodates multiple common factors driving asset returns.
- Stochastic volatility: Accounts for time-varying variance in asset returns.
- Time-varying transaction costs: Models transaction costs that can depend on market conditions.
- Predictable returns: Incorporates signals that forecast future asset returns.
The LGS methodology involves defining "exposures" (which can be non-linear functions of state variables) and then specifying trading strategies as linear functions of these exposures.
Datasets
The paper uses simulated data for its primary experiments, comparing the LGS approach against other policies in both a "characteristics model" (constant covariance) and a "factor model" (stochastic covariance).
For empirical validation, the authors use:
- CRSP data: For the 100 largest firms by market capitalization from January 1974 to March 2013.
- Ken French's data: Monthly returns of portfolios formed on momentum, short-term reversal, and long-term reversal for calibration.
Key Insights and Ideas
- Tractability through restriction: The key insight is that by restricting the strategy space to LGS, a complex problem becomes tractable, yielding a closed-form solution. This is a powerful technique for dealing with otherwise intractable optimization problems.
- LGS as a flexible framework: The LGS approach is highly flexible, allowing for the inclusion of various information sources (exposures) and complex return dynamics. The performance of the LGS depends on how well the chosen exposures span the space of optimal strategies.
- Factor models are challenging: The paper highlights that standard linear-quadratic approaches (like Litterman & Pedersen) struggle in factor-based environments where covariance matrices are stochastic. The LGS approach, by explicitly handling time-varying covariances, performs significantly better in these settings.
- Dynamic vs. Myopic: Dynamic strategies (like LGS and GP-R) that account for future return dynamics outperform myopic strategies, especially when transaction costs are present and returns are predictable.
Key Results
- Outperformance in factor models: In factor-based simulations, the Best Linear (BL) policy, an implementation of LGS, significantly outperforms other strategies, including the Litterman-Pedersen (GP) approach, by up to 14%.
- Robustness to transaction costs: The LGS approach demonstrates strong performance across various transaction cost levels, consistently outperforming myopic strategies.
- Short-term reversal strategy: An empirical application to a short-term reversal strategy shows that the LGS (specifically, the Restricted Best Linear policy) outperforms a myopic strategy, yielding higher objective values and Sharpe ratios.
- GP-R limitations: The Gârleanu & Pedersen policy with Resolve (GP-R) performs well but is suboptimal in factor models because it doesn't fully capture the dynamics of changing covariance matrices.
Limitations and Future Work
- LGS optimality: The LGS solution is optimal within the chosen set of strategies. The quality of the solution depends on selecting exposures that "span" the globally optimal solution.
- Data requirements: The LGS approach requires specifying the return generating process and estimating parameters, unlike some other methods that might rely solely on empirical performance.
- Future work: The authors suggest investigating higher transaction cost levels and exploring the dependence of transaction costs on market volatility. They also note that the LGS framework can be extended to non-zero risk-free rates and more general objective functions.
Summary updated