Will AI Turn EA into the "Holy Grail"—What We Learned from Actually Implementing It

This is an episode in the series. Recently, there has been an increasing number of EA and indicators that claim to be “AI-equipped.” When you hear phrases like “AI predicts the market” or “AI determines the optimal timing,” it’s natural to feel hopeful. In the EA world there has long been the term “holy grail (a method that always wins),” but can AI become such a holy grail?

This time, from the perspective of someone actually testing USDJPY Ultimate with an AI mechanism, I will honestly share what I feel.

What I want to tell you in advance

In this series I haven’t touched on it, but actually the USDJPY Ultimate code already includes experimental AI features. However, by default these are OFF and currently do not affect the performance of the version being sold. They are features in the development/validation stage. With this premise, I will tell you what I found when actually trying it.

What people often imagine when they hear “AI-equipped”

In the past few years, if you ask AI a question, it thinks according to the situation and returns what seems best at that moment. If you consult a chat-style AI, it understands your situation, compares several options, and provides an answer suitable for the moment.

When many people see the phrase “AI-equipped EA,” I think they unconsciously imagine something close to this. It would read the current chart, the current news, the current market mood, and AI would make the best decision each time—like a smart adviser.

The AI model embedded in EAs is a so-called “fixed calculation” that has already learned from past data. It does not rethink on the fly or understand context. The inputs are just a few numerically represented features such as ADX values, SRSI state, and Bollinger Band width—features that have been introduced in previous articles. It isn’t reading news or sensing the overall market atmosphere. What comes out is not a dialogue but a single number.

In other words, in reality it is closer to “another, somewhat complex indicator” than a “wise adviser.” This certainly doesn’t mean it’s a disappointment, but it shows there is a considerable distance between the common image of AI and the actual AI inside EAs. With this premise, I will share what I actually tested.

Three approaches I tested

Saying “embed AI in an EA” is not a single one-size-fits-all approach. In practice, I examined and implemented three major approaches.

In summary, the third approach is currently considered the most solid. The reason is simple: ① and ② challenge AI with the highly difficult task of predicting the future of the entire market, which even human discretionary traders find very hard. In contrast, ③ narrows the scope to a much more limited and manageable problem—selecting from already filtered signals and removing the weaker ones. The key lesson is to break problems into forms that AI can handle, rather than throwing a hard problem at AI all at once.

Why AI will not become the “holy grail”

Throughout this series, I have repeatedly emphasized avoiding overfitting to past data. AI models have far more internal parameters than standard rule-based filters. High expressive power means, conversely, there is a greater chance of fitting the past data by chance.

In fact, even when the learning data period shows excellent predictive accuracy, testing on another unused period can cause the accuracy to drop significantly. This is a part that I feel requires even more caution in EA development using AI than in normal rule-based validation. Strictly separate learning data from validation data and observe how it behaves on periods not used for learning. Judging that “AI performed well” without this step is dangerous.

Still, I believe there is room for use

I have been cautious here, but I am not fully denying AI. It is a domain where AI can excel, such as discovering patterns that the human mind has created features from—like ADX/MA/SRSI values and Bollinger Band states—that indicate when certain combinations are empirically unfavorable.

What matters is not to make AI the main actor or to design a system where AI’s judgment is everything. Instead, I think AI should be layered as an auxiliary support on top of the existing foundations of SRSI, the four filters, and money management. If the foundation is not solid, entrusting decisions solely to AI would create a black box that no one can explain. That is the exact opposite of the “design that hides nothing” that I have consistently advocated in this series.