This paper introduces a framework for enhancing cross-sectional systematic strategies by incorporating Learning to Rank (LTR) algorithms, which explicitly optimize for ranking accuracy rather than pointwise prediction error. Using cross-sectional momentum as a demonstrative case study, the authors compare four LTR models (RankNet, LambdaMART, ListNet, ListMLE) against classical heuristics (Jegadeesh & Titman, Baz et al.) and a regress-then-rank MLP baseline. All models use the same set of price-based features (cumulative returns, volatility-normalized returns, MACD indicators) and are evaluated on US equities from 1980 to 2019 with monthly rebalancing and volatility scaling to 15% target.

The results show that LTR models dramatically outperform all benchmarks. LambdaMART achieves the highest Sharpe ratio of 2.156, compared to 0.551 for classical momentum and 0.696 for the MACD-based strategy, representing approximately a threefold improvement. LTR models also exhibit lower drawdowns, higher Sortino and Calmar ratios, and better ranking metrics (NDCG@100 and Kendall's Tau). The decile portfolio analysis confirms that LTR models produce more accurate rankings, with monotonic return patterns across deciles and larger long-short spreads. The pairwise methods (RankNet and LambdaMART) perform comparably to listwise methods, likely due to the larger effective sample size from pairwise comparisons in noisy financial data.

The proposed framework is modular and generalizable, allowing easy substitution of different LTR algorithms and feature sets. This work highlights the importance of directly optimizing for ranking in cross-sectional strategies, providing a significant advancement over traditional approaches that rely on heuristic sorting or regression-based predictions.