DFT: A Dual-branch Framework of Fluctuation and Trend for Stock Price Prediction

Chengqi Dong, Zhiyuan Cao, S. Kevin Zhou, Jia LiuarXiv 2024 · DOI · arXiv

Predicting stock market movements is a notoriously difficult task, often likened to finding a signal amidst a hurricane. The challenge lies in the dual nature of stock data: it is a composite of stable, long-term economic trends and chaotic, high-frequency fluctuations driven by investor sentiment. Traditional deep learning models often struggle because they treat these distinct signals as a single, unified input, leading to "blurred" feature representations.

In the paper "DFT: A Dual-branch Framework of Fluctuation and Trend for Stock Price Prediction," researchers introduce a novel architecture designed to crack this code through explicit signal decomposition and a hybrid modeling approach.

The Core Challenge: Volatility and Causality

Stock price prediction faces two primary hurdles. First is the Volatility Trap: models that ignore the distinction between fundamental-driven trends and sentiment-driven fluctuations fail to capture the nuanced dynamics of the market. Second is Temporal Causality: many modern architectures, like standard Transformers, can inadvertently allow future information to leak into the past during training or fail to account for the dynamic, time-evolving relationships between different stocks.

Methodology: The DFT Framework

The proposed DFT (Dual-branch Framework of Fluctuation and Trend) addresses these issues through a structured three-step process:

  1. Decomposition: The model uses an Average Pooling (AvgPool) module to smooth raw data and extract the "Trend" component. The "Fluctuation" component is then isolated as the residual noise (original signal minus trend).
  2. Dual-Branch Strategy:
  3. Fluctuation Branch: Models time correlation first, then stock correlation. Since short-term volatility is often unique to individual stocks, the model learns the "rhythm" of a single stock before relating it to peers.
  4. Trend Branch: Models stock correlation first. Because trends are often sector-wide or market-wide, the model analyzes how groups of stocks move together before looking at individual time series.
  5. RWKV Integration: Instead of standard Transformers, DFT utilizes the RWKV (Receptance Weighted Key Value) architecture. RWKV combines the parallel training efficiency of Transformers with the linear complexity and strict temporal causality of Recurrent Neural Networks (RNNs), ensuring the model never "looks ahead" in the time series.
The DFT architecture decomposes stock data into trend and fluctuation branches, processing them with specialized modeling orders.

Strategic Insights: Why Order Matters

A key finding of the research is that the order of modeling—Time correlation vs. Stock correlation—is critical. In the fluctuation branch, reversing this order led to a significant performance drop. This confirms that short-term volatility is a "bottom-up" trait inherent to individual stocks, whereas long-term trends are "top-down" market phenomena. By exploiting these specific characteristics, DFT captures price-volume patterns that traditional "one-size-fits-all" models miss.

Empirical Results and Market Resilience

The researchers evaluated DFT against major benchmarks including the CSI300, CSI800, and S&P 500, using the Alpha158 indicator set. The quantitative results were exceptional:

  • Ranking Accuracy: On the CSI300, DFT achieved an Information Coefficient (IC) of 0.143, a nearly 3x improvement over the previous state-of-the-art model, MASTER (0.055).
  • Portfolio Performance: In backtesting a "Top30" trading strategy, DFT delivered an Annual Return (AR) of 80% on the CSI300, far outstripping competitors which typically ranged between 4% and 20%.
  • Resilience: Perhaps most importantly, DFT demonstrated remarkable resilience. During market downturns where benchmarks and other models declined, DFT maintained a positive, upward-sloping cumulative return curve.
Cumulative return plots for CSI800 and S&P500 show DFT significantly outperforming all baseline models and the market benchmark.

Conclusion and Future Directions

The DFT framework represents a significant step forward in financial time-series forecasting by proving that better decomposition is as important as more data. While highly effective on daily return ratios, future work will explore applying this dual-branch approach to high-frequency trading (HFT) and incorporating market-specific "hard rules" (like price limits) directly into the architecture. By enforcing strict causality and respecting the different natures of trends and fluctuations, DFT provides a robust blueprint for navigating the complexities of the global stock market.

Paper Citation: Dong, C., Cao, Z., Zhou, S. K., & Liu, J. (2024). DFT: A Dual-branch Framework of Fluctuation and Trend for Stock Price Prediction. arXiv:2411.06065v1 [cs.CE].

Summary updated

BibTeX
@misc{dong2024dft,
  author = {Chengqi Dong and Zhiyuan Cao and S. Kevin Zhou and Jia Liu},
  title = {{DFT: A Dual-branch Framework of Fluctuation and Trend for Stock Price Prediction}},
  year = {2024},
  eprint = {2411.06065},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.06065},
}
Figure