Fourier-Domain Features and Machine Learning for Maximum Drawdown Prediction in the CSI 1000 Index

Abstract

In the field of investment decision-making and risk management, maximum drawdown is a key measure of downside risk. Time-domain variables such as returns, volatility, historical drawdowns, trading volume, and turnover are relied on by existing studies, but frequency-domain structures receive less attention. This study aims to examine, for the CSI 1000 Index, whether frequency-domain features can improve machine learning forecasts of future maximum drawdown, and the data is from Eastmoney. Based on the conventional risk characteristics, three Fourier indicators are constructed, namely Frequency Energy Shift (FES), Spectral Entropy (SE) and Stock Market Spectral Synchronization (SSC). The samples are divided into training sets, validation sets, and test sets according to time. The model used for prediction is LightGBM. In drawdown observations, MAE, RMSE and Tail MAE are used to evaluate the overall prediction accuracy and performance. The baseline model only uses conventional features, and the full model adds these three frequency-domain indicators. Finally, the results of the two models are compared. Compared with the baseline model, the MAE of the full model has decreased by 12.67%, RMSE by 11.56%, and Tail MAE by 14.71%. The results show that the frequency-domain features can improve the overall prediction accuracy and provide some information not captured by traditional risk variables. This research framework provides a perspective for the prediction of future maximum drawdown and is also helpful for monitoring the downside risk of the stock market.