Discussing the article: "Reimagining Classic Strategies (Part 14): Multiple Strategy Analysis"

[MetaQuotes]

Check out the new article: Reimagining Classic Strategies (Part 14): Multiple Strategy Analysis.

In this article, we continue our exploration of building an ensemble of trading strategies and using the MT5 genetic optimizer to tune the strategy parameters. Today, we analyzed the data in Python, showing our model could better predict which strategy would outperform, achieving higher accuracy than forecasting market returns directly. However, when we tested our application with its statistical models, our performance levels fell dismally. We subsequently discovered that the genetic optimizer unfortunately favored highly correlated strategies, prompting us to revise our method to keep vote weights fixed and focus optimization on indicator settings instead.

There are many things that could go wrong when trying to employ artificial intelligence to help build trading strategies and it appears the genetic optimizer exploited the framework we provided and  selected the most correlated strategies. From a purely mathematical perspective, this can be seen as a clever move: it becomes easier for the genetic optimizer to anticipate the overall balance of the account when the dominant strategies are correlated.

Initially, I expected the genetic optimizer would assign higher weights to the most profitable strategies and smaller weights to the less profitable ones. However, given that we only had 3 strategies to chose from, and this optimization procedure was only performed once, we cannot rule out that all of this may have happened by chance. That is to say, if we repeated optimization of the vote weights using a slow and complete optimization algorithm, then maybe our optimizer would not have selected correlated strategies. 

This insight prompted me to revise the approach we use to select the optimal settings for our strategies. It appears we should initially keep all the weights of each vote fixed to one. This forces the genetic optimizer to focus solely on finding the most profitable settings for each indicator we use. As we shall see through our journey together, this revised approach proves to be better than our initial plan. When two correlated strategies are used for multiple strategy analysis, no real progress is made, therefore, we have learned a better way of framing the objective problem of multiple strategy analysis: "How best can we select multiple strategies that have uncorrelated returns and maximize the profitability of our account?". 

Author: Gamuchirai Zororo Ndawana