AI Learning-to-Rank for Chinese Large-Cap Index Enhancement

Summary

This article presents a supervised learning-to-rank approach to stock selection. It defines a universe using the constituents of the SSE 50, labels stocks by discretizing their subsequent five-day returns into 20 ordered classes, and trains on data from 2011 through 2015 before evaluating from 2016 to early 2018. Features include relative returns against the benchmark, with custom expressions proposed as a way to create additional features.

The strategy buys higher-ranked constituents and assigns larger weights to stronger-ranked names. It refreshes part of the portfolio daily, rather than rebalancing the entire portfolio monthly, and compares the resulting return path with the SSE 50. The article reports a 52.07% total strategy return versus 29.31% for the benchmark, plus annualized return, Sharpe ratio, and maximum drawdown figures. It also attributes much of the observed risk exposure to value and size. These are historical backtest results from the stated test interval; the document gives no evidence that the performance generalizes beyond that sample and notes that the page is an outdated version.

Key ideas

• Learning-to-rank can order stocks by expected return category within a defined index universe.
• The example labels future five-day returns in 20 ordered classes and uses a time-separated training and test period.
• Relative performance against the benchmark is included as a feature, and custom expressions can create derived features.
• The portfolio weights higher-ranked stocks more heavily and adjusts only part of its holdings each day.
• The reported index outperformance is based on a historical backtest and is concentrated in value and size exposures.

Tags

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