ListFold Ranking Loss for Long-Short Stock Portfolios

Summary

This article proposes ListFold, a listwise learning-to-rank loss for factor-based long-short equity strategies. Its central design treats the top and bottom of a predicted stock ranking as equally important, so one coherent ordering can select both long and short positions. The loss has translation invariance and is presented with a probabilistic interpretation. Depending on its transformation, the article argues that it aligns with binary classification loss or permutation-level zero-one loss.

The empirical study uses Chinese A-share data with 68 factors and compares a neural network trained with ListFold against value prediction, ListMLE, and a two-model ListMLE variant. Portfolios go long the top-ranked stocks and short either the bottom-ranked stocks or the market average. The authors report stronger out-of-sample results for ListFold, especially its exponential variant, and find that information coefficient tracks portfolio performance better than NDCG-style measures. The evidence is a historical backtest; the article notes turnover, transaction costs, sensitivity to training batch size, and the need for further theoretical work.

Key ideas

• ListFold is designed to give the top and bottom of a stock ranking equal importance.
• A single coherent ranking can determine both long and short positions.
• The study compares ListFold with value prediction and alternative ranking losses on Chinese A-share data.
• The authors report better backtest performance for ListFold, especially its exponential variant.
• The article identifies turnover, training choices, and limited theoretical analysis as caveats.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.