Training MSFformer for Multiscale Financial Time-Series Forecasting

Summary

This article describes how to train and evaluate an MSFformer-based environmental state encoder for market time-series forecasting. It combines batch normalization, a feature extraction module that builds multiscale features, and stacked pyramid attention layers. Convolutional layers then produce forecasts for each input series, with a final denormalization step restoring the original scale.

The design uses normalized raw market inputs, separate forecast trajectories for each feature, and a bounded output activation intended to reduce the influence of outliers. The article evaluates the trained model in the MetaTrader 5 Strategy Tester. The reported test includes seven transactions, with a profit factor of 1.14 and a profitable share of nearly 46%; gains occurred early, followed by flat performance and a late drawdown. This is a small test sample over a limited period, so the positive result does not establish robustness. The author suggests further training may be needed.

Key ideas

• Batch normalization is used to prepare raw market inputs for the encoder.
• CSCM extracts features at multiple temporal scales before Skip-PAM attention layers model dependencies.
• Convolutional layers map each input feature series to its own forecast trajectory.
• The reported Strategy Tester result was positive but based on few trades and included a late drawdown.

Tags

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