Enhancing Long–Short Portfolios: A Refined Approach Using Learn-to-Rank Algorithms

This article delves into the challenges posed by ranking biases inherent to traditional learn-to-rank loss functions, particularly focusing on their impact on the construction of long–short portfolios. Through the analysis of synthetic data, the authors uncover inherent biases in these methods, particularly detrimental for long–short portfolios where equal importance lies with top- and bottom-ranked assets. To address these challenges, the authors propose enhanced versions of learn-to-rank loss functions—ListNet-Fold, ListMLE-weighted, and ListFold-weighted. These adaptations, tailored for long–short strategies, draw inspiration from pairwise approaches and adjust weighting mechanisms. Empirical results using a real-world dataset sourced from the China A-share market consistently reveal enhancements in ranking metrics, notably improving accuracy in ranking extreme assets, which are more traded in long–short portfolios. Furthermore, financial performance metrics validate the efficacy of these methods, demonstrating enhanced risk-adjusted returns, profitability, and robustness across varying numbers of assets included in the long–short strategy. This research offers valuable insights and practical remedies for mitigating biases in learn-to-rank algorithms, presenting promising tools for constructing long–short portfolios.