What question did this study set out to answer?

The aim is to improve financial time series forecasting through advanced modeling techniques that capture complex temporal dependencies.

April 23, 2026Open Access

MSTFNet: Multi-Scale Temporal Fusion Network with Frequency-Enhanced Attention for Financial Time Series Forecasting

Key Points

The aim is to improve financial time series forecasting through advanced modeling techniques that capture complex temporal dependencies.
Developed MSTFNet utilizing dilated causal convolutions and frequency-enhanced sparse attention.
Applied the model to three public financial datasets: S&P 500, CSI 300, and NASDAQ Composite.
Conducted extensive experiments using various level-error metrics and directional accuracy.
MSTFNet outperformed traditional models, achieving a directional accuracy of 56.3% for the S&P 500.
Reduced MAE by 13.6% compared to the best baseline (iTransformer) for the S&P 500.
Established random-walk as a competitive benchmark in financial forecasting.

Abstract

Financial time series forecasting remains a persistent challenge due to the non-stationary nature, inherent noise, and multi-scale temporal dependencies present in market data. This paper presents MSTFNet, a multi-scale temporal fusion network that combines dilated causal convolutions with a frequency-enhanced sparse attention mechanism for improved financial prediction. The proposed architecture consists of three core components: a multi-scale dilated causal convolution module that extracts temporal patterns across different time horizons through parallel convolutional branches with varying dilation rates, a frequency-enhanced sparse attention mechanism that leverages Fast Fourier Transform to identify dominant periodic components and modulate attention weights accordingly, and an adaptive scale fusion gate that learns to dynamically combine representations from multiple temporal scales. Extensive experiments conducted on three public financial datasets (S&P 500, CSI 300, and NASDAQ Composite) spanning the period from January 2015 to December 2024 show two key results. First, consistent with near-efficient markets, the random-walk benchmark (yᵗ+1=yt) outperforms all the data-driven models on level-error metrics (MAE, RMSE, MAPE, and R2), establishing the martingale as the binding lower bound on point-prediction error. Second, MSTFNet achieves the highest directional accuracy (DA) across all three indices—56. 3% on the S&P 500 versus 50. 0% for the martingale—representing a 6. 3 percentage-point improvement that generates positive pre-cost returns in a trading strategy backtest. Among the eight data-driven baselines (LSTM, GRU, TCN, Transformer, Autoformer, FEDformer, PatchTST, and iTransformer), MSTFNet also achieves the lowest MAE, reducing it by 13. 6% relative to the strongest data-driven baseline (iTransformer) on the S&P 500. These results confirm that integrating multi-scale temporal modeling with frequency-domain guidance extracts a real, if modest, directional signal from financial time series.

MSTFNet: Multi-Scale Temporal Fusion Network with Frequency-Enhanced Attention for Financial Time Series Forecasting

Key Points

Abstract

Cite This Study