Using Transformer Models to Predict Five-Day Returns in Chinese Stocks

Summary

This study outlines an experiment applying a Transformer model to short-horizon stock selection in China’s A-share market. Its target is each stock’s return over the next five trading days, using daily market data from 2015 through 2021. The inputs begin with seven price and trading variables, from which the authors construct 98 features using rolling statistics, correlations, and lagged observations. Each sample contains a five-day sequence of features.

The described preprocessing clips target returns at the cross-sectional 1st and 99th percentiles and standardizes them. Features are cross-sectionally standardized, missing values are filled, and values are clipped to a bounded range. Hyperparameters are tuned on an earlier 2011–2013 period and then carried into rolling training. The supplied text stops at the experiment setup: it does not report model architecture details, comparison results, portfolio construction, transaction costs, or out-of-sample performance. The setup therefore describes a research design, not evidence that the model produced profitable selections.

Key ideas

• The experiment uses a Transformer to forecast five-day stock returns in China’s A-share market.
• Seven basic market variables are transformed into 98 features using rolling statistics, correlations, and lags.
• The feature sequence spans five days, and both labels and features receive cross-sectional preprocessing.
• Hyperparameters are tuned on an earlier period before being applied in rolling training.
• The provided material contains no model results or evidence of investment performance.

Tags

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