Discussing the article: "Neural Networks in Trading: Hierarchical Dual-Tower Transformer (Final Part)"

 

Check out the new article: Neural Networks in Trading: Hierarchical Dual-Tower Transformer (Final Part).

We continue to build the Hidformer hierarchical dual-tower transformer model designed for analyzing and forecasting complex multivariate time series. In this article, we will bring the work we started earlier to its logical conclusion — we will test the model on real historical data.

We have completed substantial work in implementing our interpretation of the approaches proposed by the Hidformer authors. We now arrive at the crucial stage: evaluating the effectiveness of our solutions on real historical data. In our implementation, we borrowed extensively from the MacroHFT framework. Therefore, it's logical to compare the new model performance with it. So, we train the new model on the training dataset previously compiled for training the MacroHFT-based implementation.

That training dataset was collected from historical data for the entire year of 2024 for the EURUSD currency pair on the M1 timeframe. All indicator parameters were set to their default values.

The same Expert Advisors are used for training and testing the model. Testing was conducted on historical data from January 2025, maintaining all other parameters. The test results are presented below.

The results show that the model achieved profit on historical data outside the training dataset. Overall, during the calendar month, the model executed 29 trades. This makes slightly more than one trade per trading day, which is not enough for high-frequency trading. Over 60% of trades were profitable. The average profitable trade is 60% higher than the average losing trade.


Author: Dmitriy Gizlyk