[EA制作] Logic Exploration ①

Hello. This is 2pay!

Today I intend to pioneer some logic.

The theme of the logic is "NY Open".

It is known that just after the NY stock market opens, positions from the London side tend to unwind. (I don't know)

I've heard about it, but I've never seen an implementation example as an EA, and I want to try making one to verify whether I can use it myself.

Rules

EURUSD M5

① Get the price movement direction for the 3 hours before NY open

② At NY open, enter in the opposite direction to (①)

③ Hold for 30 minutes and close

When I first create a logic, I always test with such simple rules.

If you have one or two core conditions that truly capture the essence, they can function for more than 20 years, so the key is not to pile on many conditions and to discard anything that isn’t usable as it becomes unnecessary.

I’m not very confident in these conditions this time, but let’s run a light 10-year pass.

The final profit and loss almost returned to break-even. It seems there was no edge (i.e., random).

Try to improve

If this were just my whim, I’d drop the idea, but knowledge of NY Open tendencies is recognized by hedge funds and other active investors.

I think it was even discussed on sites like arXiv where papers are published.

In other words, there is a high possibility that my approach has issues.

I thought for about 3 seconds, but I couldn’t come up with any solution, so I’ll ask ChatGPT (Chapiemon) for help.

Below are the improvement requests I received from Chapiemon.

・Measure the “strength” of European trends.・Check whether stops were wiped out before NY open.
・Entries are too early. Defend against the initial move algorithm.
・European trend capture range is too short. Look at Europe as a whole.
・TP/SL are wide. Don’t move so much on the 30-minute rebound.
・Check the day of the week.

・Change the market. GBPUSD is better.
・Avoid on events days.
・Also, since 2015 it has weakened.

Those are wonderful words.

I will incorporate this into the code and run it again.

Wouldn’t you say it went pretty well?

In the last 10 years there were only 2 years where I underperformed. Most years show a steady increase.

Listening to human (AI) advice is important, isn’t it?

This is still plain, not yet optimized for currency, scale, or parameters. From here, optimize within reasonable bounds and add filters to make a sufficiently practical system.

The idea that it weakens after 2015 is partly dependent on the logic, so I’ll just try building it as a project. (As a rule of thumb, structural distortions are not easy to eliminate.)

EA improvements are good, but since I can learn about institutional investors and the financial industry, I chat with AI from morning to night.

You hear that even if you ask AI to listen and create, it won’t produce something sensible, but that depends on the prompts you give.

For example, if you ask AI to “Create a winning EA using RSI and Bollinger Bands.”

AI will create an EA, but it won’t be winning. The reason is that the indicators themselves do not have a built-in edge.

The answer to the question “If RSI falls below 30 and then breaks above 30, what will happen to the price after this?” is “Random (will move in either direction with 50% probability).”

(No matter how you use it, the result remains random.) This is also the gambler’s fallacy.

With this basic understanding, you won’t impose the futile constraint “use RSI” on AI.

It's like telling AI, “Please devise a strategy to defeat the demon king using the starter sword and shield.” And if you test that strategy and it instantly collapses, people will criticize, “AI is useless, garbage.”

From AI’s perspective, there is no such unfairness.

I’m not sure where this discussion is heading, so I’ll summarize roughly: to obtain proper answers from AI, the user must possess appropriate knowledge. If you throw garbage, you’ll get garbage back.

Though it may annoy colleagues, I’ll say a little.

You might wonder what to use if RSI doesn’t work. Open, High, Low, Close, Time (OHLCT{Open[], High[], Low[], Close[], Time[]}) is sufficient.

I’ll twist and turn this based on financial engineering and market structure.

Autocorrelation and regression are fine, or you can include other currencies as well.

Many top traders say they don’t use indicators and only use candles. They probably learned that from experience, and it’s been statistically shown that similar results are obtained.

That’s all for now.

Thank you for reading to the end.