AFD-Net: a robust exchange rate forecasting framework integrating frequency-domain decomposition and attention...

Area of Science:

• Quantitative Finance
• Machine Learning
• Time Series Analysis

Background:

• Exchange rate forecasting is challenging due to non-linearity and non-stationarity.
• Existing models struggle with complex temporal dynamics in currency markets.

Purpose of the Study:

• Introduce AFD-Net, a novel deep learning framework for enhanced exchange rate prediction.
• Synergistically integrate frequency-domain decomposition, attention mechanisms, and linear time series modeling.

Main Methods:

• Employ Feature-wise Adaptive Normalization (FAN) for frequency decomposition and stationarity.
• Utilize an attention module to capture salient temporal dependencies.
• Leverage a DLinear backbone for efficient trend and seasonal component prediction.

Main Results:

• AFD-Net consistently outperforms seven state-of-the-art models on Exchange-Rate and IMF datasets.
• Demonstrates superior performance in long-term trend modeling.
• Ablation studies confirm the contributions of FAN and attention components.

Conclusions:

• AFD-Net offers enhanced accuracy and robustness in financial time series forecasting.
• Frequency-aware architectures show significant potential for improving prediction models.
• The FFT-based design of FAN provides substantial performance gains with manageable computational overhead.