Decomposition-Enhanced Network for financial time series forecasting

Area of Science:

• Quantitative Finance
• Machine Learning
• Time Series Analysis

Background:

• Financial futures markets exhibit extreme non-stationarity, high noise, and multi-timescale coupling, challenging traditional forecasting models.
• Existing models struggle to differentiate localized shocks from global trends due to incompatible inductive biases.

Purpose of the Study:

• To propose a novel Decomposition-Enhanced Network (DENet) for accurate time series forecasting in financial futures markets.
• To enhance algorithmic trading strategies by integrating advanced structural modeling with actionable predictions.

Main Methods:

• DENet employs a multi-stream architecture: a main path for stable trends (moving averages, dual-path linear projections), Auxiliary Stream I for local patterns (depthwise separable convolutions), and Auxiliary Stream II for high-frequency dynamics (nonlinear autoregressive mapping).
• An adaptive fusion mechanism integrates these streams, balancing global robustness and local sensitivity.
• DENet's predictions were integrated into the R-Breaker trading strategy with parameter switching and dynamic position sizing.

Main Results:

• DENet significantly outperforms state-of-the-art benchmarks on real-world iron ore futures data.
• Achieved an average 10.15% RMSE reduction in daily forecasting and a 22.57% MAE reduction for 5-min, 12-step horizons.
• The integrated trading strategy yielded an average 7.6 percentage point higher annualized return compared to the baseline strategy.

Conclusions:

• DENet effectively addresses the challenges of non-stationarity and multi-timescale dynamics in financial futures forecasting.
• The model provides a robust framework for bridging advanced structural modeling with practical algorithmic trading applications.
• DENet demonstrates superior performance in both forecasting accuracy and enhancing trading strategy profitability.