Discussing the article: "Neural Networks in Trading: Hierarchical Feature Learning for Point Clouds"

[MetaQuotes]

Check out the new article: Neural Networks in Trading: Hierarchical Feature Learning for Point Clouds.

We continue to study algorithms for extracting features from a point cloud. In this article, we will get acquainted with the mechanisms for increasing the efficiency of the PointNet method.

As mentioned earlier, our new model differs from the previous one by only a single layer. Furthermore, this new layer is merely an improved version of our prior work. This makes it particularly interesting to compare the performance of both models. To ensure a fair comparison, we will train both models on the exact same dataset used in the previous experiment.

I always emphasize that updating the training dataset periodically is crucial for achieving optimal model performance. Keeping the dataset aligned with the Actor's current policy ensures a more accurate evaluation of its actions, leading to policy refinements. However, in this case, I couldn’t resist the opportunity to compare two similar approaches and assess the effectiveness of a hierarchical method. In our previous article, we successfully trained an actor policy that was capable of generating profit. We expect the new model to perform at least as well.

After training, our new model successfully learned a profitable policy, achieving positive returns on both training and test datasets. The test results for the new model are presented below.

I must admit that comparing the results of both models is quite challenging. Over the test period, both models generated nearly the same profit. Drawdown deviations in both balance and equity remain within a negligible margin of error. However, the new model executed fewer trades, leading to a slight increase in the profit factor.

That being said, the low number of trades executed by both models does not allow us to draw definitive conclusions about their long-term performance.

Author: Dmitriy Gizlyk