Testing Low-Pass Filtering and Seasonal Training for Forex LSTM Models

Summary

The article compares an EURUSD hourly LSTM model trained on unfiltered prices with versions trained after low-pass filtering, then explores training on February data concatenated across years. It also varies the filter settings and prediction horizon across chart periods, describing the use of a MinMax scaler and suggesting a maximum holding time as a possible EA adjustment.

For one February test period, the reported filtered version had lower RMSE and higher R² than the unfiltered model, while another filter configuration performed worse on those metrics. The seasonal model had a lower R² than the filtered comparison, yet the author reports fewer negative Sharpe outcomes in an optimization table. These results suggest filtering and seasonal training may affect both prediction measures and trading robustness, but the evidence is limited to the author's setup and reported comparisons. The article does not provide a broad out-of-sample evaluation or establish that the approach generalizes to other symbols or periods.

Key ideas

• The article compares EURUSD LSTM models trained with and without a low-pass filter.
• A filter configuration that improved reported prediction metrics did not make all tested configurations better.
• The seasonal model was trained on February observations concatenated across multiple years.
• The author reports fewer negative Sharpe outcomes for the seasonal model despite its lower R².
• The article suggests tuning the prediction horizon and limiting how long a trade remains open.

Tags

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