Time‐Frequency Market‐State Learning for Execution‐Aware Directional Prediction: Evidence From the CSI 300 Index

ABSTRACT Financial markets can be viewed as multiscale information systems in which short‐lived disturbances, recurrent temporal patterns, liquidity conditions, and execution constraints jointly shape index price movements. This study develops an empirical‐wavelet‐and‐Fourier time‐frequency (EFT) feature block integrated with a ConvTransformer architecture, hereafter referred to as the EFT‐ConvTransformer, for one‐day‐ahead directional prediction of the CSI 300 Index. The EFT block combines empirical wavelet transform (EWT)‐based high‐frequency representations with fast Fourier transform (FFT)‐based short‐window periodic‐energy descriptors, while the ConvTransformer architecture incorporates convolutional local‐pattern extraction and Transformer‐based attention. Daily observations of CSI 300 prices, trading volume, trading value, and technical indicators from January 2005 to September 2024 are used as forecasting inputs, whereas matched minute‐level observations are reserved for separate market‐state validation and delayed‐execution evaluation. During the out‐of‐sample test period from 2023 to September 2024, the EFT‐ConvTransformer achieves an area under the receiver operating characteristic curve (AUC) of 0.674, an accuracy of 61.8%, and a Matthews correlation coefficient (MCC) of 0.236, representing the strongest overall predictive performance among the benchmarks considered. Under a delayed closing‐period volume‐weighted average price (VWAP) execution convention and notional exchange‐traded fund (ETF)‐style transaction‐cost assumptions, the signal‐based timing illustration yields an annualized return of 11.4%, a Sharpe ratio of 0.88, and a maximum drawdown of 15.4%. The market‐state validation results indicate that EWT high‐frequency energy is primarily associated with realized volatility and downside price adjustment, whereas horizon‐aligned FFT periodic energy is associated with recurrent calendar‐related states. Overall, the findings suggest that time‐frequency representations can improve out‐of‐sample short‐horizon index‐direction forecasting in the CSI 300 setting while providing economically interpretable, non‐causal diagnostic evidence on underlying market states.