SCNN Time-Series Decomposition and Its EURUSD Backtest

Summary

The article concludes an implementation of a Seasonal Convolutional Neural Network for financial time-series forecasting. The model separates long-term, seasonal, and short-term components through normalization and seasonal transformations, then combines those representations with spatial normalization, attention, projections, and convolutional fusion. The architecture is integrated into an Actor-Critic framework. The discussion emphasizes interpretability and explains how normalized series and their statistical parameters are assembled while limiting redundant data expansion to conserve memory.

The reported test uses EURUSD data from January through March 2025. The author says the model generated profit and balanced forecasts, but also experienced deep drawdowns during sharp March market changes outside the training sample. This example highlights the gap between fitting familiar conditions and handling regime shifts. The article regards the implementation and backtest as preliminary evidence, and calls for more training and additional risk controls; a short test on one currency pair cannot establish general robustness or live-trading performance.

Key ideas

• SCNN decomposes a time series into long-term, seasonal, and short-term components before forecasting.
• The encoder normalizes these components and combines them with spatially informed representations and statistical parameters.
• The model is integrated into an Actor-Critic architecture for trading experiments.
• A EURUSD backtest covering January through March 2025 reportedly showed profit alongside deep drawdowns during sharp market changes.
• The brief single-pair test leaves robustness uncertain and points to further training and risk controls.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.