RM-CVaR: Regularized Multiple β-CVaR Portfolio

Kei Nakagawa, Shuhei Noma, Masaya AbeIJCAI 2020 · DOI

TL;DR

This paper introduces RM-CVaR, a novel portfolio optimization strategy that addresses the sensitivity of Conditional Value-at-Risk (CVaR) to its probability parameter ($\beta$). By considering multiple $\beta$ values simultaneously and incorporating regularization for transaction costs, RM-CVaR demonstrates superior risk-adjusted returns and lower maximum drawdown compared to existing methods.

What Problem Does This Paper Address?

Traditional portfolio optimization often relies on measures like variance, which can be insensitive to extreme losses. While Value-at-Risk (VaR) and its coherent counterpart, Conditional Value-at-Risk (CVaR), offer improvements by focusing on tail risk, CVaR optimization can be highly sensitive to the chosen probability level ($\beta$). Small changes in $\beta$ can lead to drastically different portfolio allocations, making the optimization unstable. Furthermore, many existing portfolio optimization models overlook the impact of transaction costs, which can significantly erode returns in real-world trading scenarios. This paper aims to bridge this gap by proposing a robust CVaR-based portfolio optimization method that accounts for multiple risk sensitivities and transaction costs.

Methods

The core of this paper is the introduction of the RM-CVaR (Regularized Multiple $\beta$-CVaR) Portfolio. The authors extend the standard CVaR optimization by:

  • Multiple $\beta$ Levels: Instead of optimizing for a single $\beta$, RM-CVaR considers a set of $\beta$ values ($K$ levels: $\beta_1, \dots, \beta_K$). This is achieved by formulating a problem that minimizes the maximum CVaR across these different probability levels, effectively hedging against the sensitivity to a single $\beta$.
  • Linear Programming Formulation: The proposed RM-CVaR optimization problem is shown to be equivalent to a linear programming problem, making it computationally tractable and efficient to solve. This builds upon the existing linear programming formulation for single $\beta$-CVaR optimization.
  • L1-Regularization for Transaction Costs: To address transaction costs and portfolio turnover, an L1-regularization term is incorporated into the objective function. This encourages sparsity in the portfolio weights, leading to lower turnover and thus reduced trading costs. The regularization term is defined as $||w - \bar{w}||_1$, where $w$ is the current portfolio weight and $\bar{w}$ is the previous portfolio weight.

Datasets

The authors evaluate their proposed RM-CVaR strategy using two well-known academic benchmark datasets:

  • Fama and French (FF) 25 dataset: Portfolios formed based on size and book-to-market ratio.
  • Fama and French (FF) 48 dataset: Portfolios representing different industrial sectors.

Both datasets consist of monthly returns from January 1989 to December 2018. The in-sample period (January 1989 to December 2003) was used for hyperparameter tuning, and the out-of-sample period (January 2004 to December 2018) was used for performance evaluation, with portfolios rebalanced monthly.

Key Insights and Ideas

  • CVaR Sensitivity is a Real Problem: The paper empirically confirms that CVaR portfolio weights are highly sensitive to the chosen $\beta$ level, with even a 1% change in $\beta$ leading to significant shifts in portfolio allocation (average of 48% difference in weights). This highlights the need for more robust methods.
  • Multiple $\beta$ Levels Offer Stability: By considering a range of $\beta$ values, RM-CVaR mitigates the instability associated with single-point CVaR optimization, leading to more consistent and reliable portfolio construction.
  • Regularization is Crucial for Real-World Performance: The inclusion of L1-regularization effectively controls portfolio turnover, which is directly linked to transaction costs. This leads to better risk-adjusted returns and lower maximum drawdowns, demonstrating the practical importance of accounting for trading frictions.
  • RM-CVaR Outperforms Existing Methods: The proposed RM-CVaR strategy, particularly with the regularization term, consistently achieves superior performance across various metrics, including risk-adjusted return (R/R) and maximum drawdown (MaxDD).

Key Results

  • Weight Sensitivity: Table 1 shows that a 1% difference in $\beta$ can change portfolio weights by an average of 48% for FF25 and 39% for FF48 datasets.
  • Performance Comparison: In out-of-sample tests on both FF25 and FF48 datasets, RM-CVaR with the regularization term ($\lambda > 0$) achieved the highest risk-adjusted return (R/R) and the lowest maximum drawdown (MaxDD) among all tested portfolios.
  • Transaction Cost Impact: RM-CVaR without regularization ($\lambda = 0$) showed high annualized returns but also very high turnover (TO). Introducing regularization significantly suppressed TO while maintaining excellent performance.
  • Visual Evidence: Figures 2 and 3 illustrate that RM-CVaR avoids large drawdowns, especially evident in the FF48 dataset, where it shows visible margins of outperformance.

Limitations and Future Work

  • Transaction Cost Model: The paper uses L1-regularization as a proxy for transaction costs. While effective, a more explicit modeling of transaction costs could be explored.
  • Hyperparameter Tuning: The choice of $\lambda$ (regularization coefficient) and the set of $\beta$ values are crucial. The paper implements several patterns, but further research into optimal selection strategies for these parameters could be beneficial.
  • Subsampling Methods: The authors suggest incorporating subsampling methods, as explored in prior work (e.g., Shen and Wang, 2017; Shen et al., 2019), as a future direction to further enhance the robustness and efficiency of their approach.
  • Machine Learning Applications: The paper notes the connection between portfolio optimization and machine learning and suggests applying their method to machine learning algorithms as a future task.

Summary updated

BibTeX
@inproceedings{nakagawa2020rm,
  author = {Kei Nakagawa and Shuhei Noma and Masaya Abe},
  title = {{RM-CVaR: Regularized Multiple β-CVaR Portfolio}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/629},
}
Figure