MF-TimesNet: An FFT-Driven Multi-Factor Fusion Framework for Stock Directional Prediction

Stock directional prediction remains a critical yet highly challenging task in quantitative finance due to the extreme noise, non-linearity, and non-stationarity inherent in financial time-series data. Traditional linear models and isolated technical indicators generally fail to adapt to complex market regimes. To address these issues, we develop a comprehensive, end-to-end prediction framework that integrates multi-factor fusion with advanced time-series modeling. Our data pipeline aggregates hour-level Open-High-Low-Close-Volume (OHLCV) data, macroeconomic variables, sentiment indices, and company quality indicators spanning from 2019 to 2025. Stable predictive features are systematically extracted using a rolling quarterly Spearman Information Coefficient (IC) screening process. For temporal sequence modeling, we introduce MF-TimesNet, a novel architecture adapted from TimesNet. By leveraging the Fast Fourier Transform (FFT), MF-TimesNet transforms 1D multi-factor time-series data into 2D temporal spaces, which adaptively capture multi-scale patterns across both intra- and inter- period variations. Ablation experiments were conducted to determine the look-back window and the number of TimesBlock for the optimal validation Area Under the Curve (AUC). Empirical evaluations show that our MF-TimesNet consistently outperforms the CNN-LSTM baseline model, demonstrating robust trend discriminability across high-liquidity assets. Specifically, the model delivers solid directional predictions for leading technology stocks: yielding an AUC of 0.5461, Recall of 0.8612, and F1-score of 0.7282 for Alphabet (GOOGL); an AUC of 0.5439, Recall of 0.9070, and F1-score of 0.7263 for Microsoft (MSFT); and an AUC of 0.5314, Recall of 0.9518, and F1-score of 0.6990 for Apple (AAPL). These empirical findings confirm that mapping multi-scale temporal dynamics into 2D convolutional spaces effectively extracts stable alpha signals from low-signal equity markets. While the absolute AUC values are strictly bounded by the near-random walk behavior of short-term asset pricing under market efficiency, the high rRcall and balanced F1-scores demonstrate a viable, statistically significant predictive edge suitable for systematic quantitative execution.