"Second-generation EA that memorizes charts (paid information available)" (6/6) Investment AI development struggles

In this article, I will tweet about what seems to be the current mainstream to the second generation of EAs.

The paid portion contains content that could be abused, so I am not writing this with the intention of wanting to sell it in a big way.

This article is for those who are stuck in EA creation, unable to break free from curve fitting, or who want to gain new insights and need some stimulation.

In the paid article, I introduce the phenomenon of memorizing charts with concrete code, and also introduce the method of self-optimization.

The content is aimed at intermediate to advanced users.

First of all, the term “generation” of EAs is something I arbitrarily analogize to AI generations and call it that.

AI generations

First generation: Rule-based, early expert systems

Second generation: Peak era of expert systems

Third generation: Statistical machine learning era (decision trees, logistic regression)

Fourth generation: Deep learning era

Fifth generation: Generative AI era

Therefore

First-generation EAs are simple rule-based EAs. They fully reflect human intent

Second-generation EAs are created from complex processes. They reflect some human intent through optimization and other techniques, but largely rely on logic developed by computers

and are EAs from the pre-machine-learning era.

However

I believe that second-generation EAs could not surpass first-generation EAs

because they end up mass-producing EAs that simply fit past markets.

By contrast, first-generation EAs were created by humans from their experience, without computer logic, making them harder to fit.

The reason I titled this piece “Memorizing Charts” is

is that curve fitting is simply a state of having memorized past charts

The cause lies largely inthe number of parameters.

Having many parameters tends to cause the EA to memorize entry points on charts more readily.

The number of parameters equals the data capacity.

So, concretely, how can an EA memorize past charts?

And how should second-generation EAs be developed in the future?

I have pondered this.

From here, this becomes a paid article