Comparing Language Models for Equity Liquidity Factor Discovery

Summary

This report compares several Chinese language models as assistants for discovering equity liquidity factors over a specified historical period. Each model proposes data inputs and a formula built from trading volume, turnover, amount, or price. The author evaluates the resulting candidates using information coefficient and portfolio performance measures, including return, Sharpe ratio, volatility, and drawdown. One turnover-based formula combining a moving average with dispersion receives the strongest assessment among the initial candidates, leading the author to favor that model for further exploration.

The report then tests additional factors proposed with the selected model, including turnover amount variability and a price deviation scaled by volatility. Their reported results vary, and the author characterizes the latter as weak. These examples illustrate a workflow for using language models to generate factor hypotheses and then screening them quantitatively. The findings are limited: the document does not establish robust out-of-sample performance, detail transaction costs or portfolio construction, or demonstrate that the model ranking generalizes beyond this dataset and period.

Key ideas

• The exercise asks multiple language models to propose liquidity factors using historical equity data.
• Candidate formulas use turnover, trading volume, traded amount, and price behavior.
• Information coefficient and portfolio statistics are used to compare the proposed factors.
• The author prefers a turnover-based candidate that combines an average with a measure of dispersion.
• The reported comparisons are exploratory and do not establish out-of-sample robustness.

Tags

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