"What makes a winning EA simple" (4/21) Investment AI Development Struggle Journal

EA Decluttering

It is a pitfall to cause curve fitting, which is the widely known fact in the EA communityKnown fact

What is needed to avoid curve fitting is simplifying the conditions

This doesn’t just mean that coding becomes easier

It's easy for humans to understand, but hard when coded

For example, like Gotovi

Logic simple = Code simple?

What must not be misunderstood is that it is better for the logic to be simple, and that simple logic does not necessarily mean simple code

〇 Simple logic

× Simple code

With the logic of GXDX (Golden Cross, Dead Cross), you should not adjust whether MA parameters are good

This leads to a situation where the logic is complex but the code is simple, which is a mistaken thinking.

So where is it simple and where is it complex? What is good and what is bad?

Usually the only way to know is to validate through backtesting, walk-forward tests, etc.

Simple with Machine Learning

Machine learning can quantify the complexity of a logic as parameters

By quantifying, you can adjust how many parameters to increase curve fitting

For example, suppose there are many kinds of indicator values, say over 1,000

This makes the code extremely complex.

However, if the number of parameters is 1, the logic becomes very simple.

This is the correct direction of simplicity.

By the way

What can be mounted in EA is

From 1 to hundreds of millions of parameters, perhaps, but if it goes into the hundreds of millions, the data becomes heavy and cannot be used for trading

LLMs, for example, are said to be over 300 billion with GPT-3.5

At this level, with consumer PCs it is not realistically feasible at the moment. It takes too long