Discussing the article: "Neural networks made easy (Part 52): Research with optimism and distribution correction"

[MetaQuotes]

Check out the new article: Neural networks made easy (Part 52): Research with optimism and distribution correction.

As the model is trained based on the experience reproduction buffer, the current Actor policy moves further and further away from the stored examples, which reduces the efficiency of training the model as a whole. In this article, we will look at the algorithm of improving the efficiency of using samples in reinforcement learning algorithms.

As usual, we are much more interested in the efficiency of the model on new data. The generalization ability and performance of the model on unfamiliar data was tested in the strategy tester on historical data for June 2023. As we can see, the testing period immediately follows the training set. This ensures maximum homogeneity of the training and test samples. The test results are presented below.

The presented chart shows a drawdown area in the first ten days of the month. But then it is followed by a period of profitability, which lasts until the end of the month. As a result, the EA received a profit of 7.7% over the course of the month with a maximum drawdown in Equity of 5.46%. In terms of the balance, the drawdown was even smaller and did not exceed 4.87%.

The table of test results shows that during the test the EA performed trades in both directions. A total of 48 positions were opened. 54.17% of them were closed with a profit. The maximum profitable trade is more than 3 times higher than the maximum losing one. The average profitable trade is half as much as the average losing trade. In quantitative terms, on average, for every 3 profitable trades there are 2 unprofitable ones. All this gave a profit factor of 1.74 and a recovery factor of 1.41.

Author: Dmitriy Gizlyk