AI-like Chart Prediction EA Development Notes 4｜When you add averaging down, it turned out to be more remarkable than imaginable. However, the real work starts from here.
Hello, this is Tsumo.
This article isNotes on AI-style Chart Prediction EA Development, Part 4.
This time, there were quite a few major changes.
Honestly, I was a little surprised.
Until now, the AI-style chart prediction EA was tested as a single-entry type.
From historical charts, look for situations similar to the current shape.
Then, by watching the subsequent price movement, decide BUY or SELL.
However, we would not excessively optimize for each currency pair.
We would also check how it behaves across multiple currency pairs with minimal崩れ (breakdowns).
We were proceeding with this policy.
As a result of the last results, although not flashy, PF for multiple currency pairs landed roughly around 0.97–1.05.
In other words,
“A winning EA has been created”
not quite,
“A foundation worth continuing to validate as a single-entry-type has emerged.”
That was the stage.
This time, we move to the next step.
For the AI-style chart prediction EA that has begun to take shape as a single-entry type,we added a Nampin (averaging down) / Basket recovery structure.
As a result, it turned out to be more transformative than I imagined.
However, please do not misunderstand here.
Adding averaging down does not mean it will win.
Narrow the entry with AI, recover with a basket, control the collapse with MTP.
This combination showed potential.
If you’re curious about introducing MTP, please also review this.
MTP: Master【Free】
https://www.gogojungle.co.jp/tools/indicators/79103?via=usersMTP: Slave【Paid】
https://www.gogojungle.co.jp/tools/indicators/79106?via=users
Position up to now
Previously, I stopped adjusting to fit past data.
Optimize numbers for only certain currency pairs.
Raise PF over only certain periods.
Fine-tune parameters to fit the past data.
That direction is dangerous.
That is what I concluded.
So, I expanded as a single-entry-type across multiple currency pairs.
The results roughly align with the following image.
EURUSD M15.
PF is 0.99.
Net profit is slightly negative.
Maximum drawdown is 3.72%.
USDJPY M15.
PF is 1.01.
Net profit is a small positive.
Maximum drawdown is 4.01%.
GBPUSD M15.
PF is 1.04.
Net profit is positive.
Maximum drawdown is 4.09%.
AUDCAD M15.
PF is 0.97.
Net profit is negative.
Maximum drawdown is 5.15%.
EURJPY M15.
PF is 1.05.
Net profit is positive.
Maximum drawdown is 1.97%.
It wasn't flashy.
But what mattered was that it didn’t break down dramatically.
In particular, EURJPY did not collapse much even with time-split testing.
2020–2022.
2023–2024.
2025–May 2026.
In any period, PF did not collapse significantly.
This wasn’t a bad impression for a single-entry type.
However, staying there would still be far from real operation.
So, I moved on to the next experiment.
Recovery-type Martingale didn’t work well
First, I tried the recovery-type Martingale.
However, the recovery-type Martingale here is slightly different from the usual averaging-down.
It doesn’t add more positions.
After a loss, in the next AI decision entry, the lot size is increased a little.
The image is like this.
Normally 0.10 lot.
If it loses, next is 0.13 lot.
If it loses again, 0.16 lot.
If it loses again, 0.21 lot.
In other words, only when a valid AI signal appears, the next lot is increased a bit.
This isn’t a bad idea.
You don’t break the single-entry logic, and you improve recovery on the next valid entry.
But when actually tested, the effect wasn’t significant.
Reason is simple.
The single-entry logic was already designed to lose small amounts.
Lose small amounts.
Recovery steps progress.
Lot sizes rise.
But there isn’t enough distortion to recover big losses.
As a result, recovery-type Martingale had little role.
Moreover, increasing lot size unnecessarily could erode the simplicity of the single-entry logic.
This prompted a re-evaluation.
Single-entry logic is not as bad as expected.
But it’s not worth making recovery-type Martingale the main approach.
So, for the time being, I set recovery-type Martingale aside.
Compared with existing EAs
Next, I compared it with the EAs we previously used.
The comparator was a Nampin/Basket-type EA.
In both, if price moves against you, you add in the same direction.
Aim to improve the average entry price.
Recover with a one-shot pullback.
That is the classic approach of a Nampin EA.
One of the existing EAs was fairly strong.
PF is 1.68.
Win rate is 78.50%.
Relative DD is 15.83%.
Net profit is also large.
That is simply strong.
However, as a Nampin EA, it cannot be considered completely safe.
Lot sizes accumulate.
If the price moves against you, unrealized losses increase.
You need to wait for a rebound.
This structure does not change.
Another EA was the type that endures if you have ample funds.
PF is 1.71.
Net profit is also large.
But, maximum drawdown is large as well.
Capital efficiency is heavy.
In other words, the logic itself isn’t bad.
But it is an EA that endures by increasing capital.
From this comparison, the clear takeaway is that Nampin EAs have a definite strength, but their failure mode is also quite clear.
When they win, they’re strong.
But a misdesign during drawdown can be heavy.
How to control this?
This connects to the theme on the MTP side as well.
Added Nampin to AI-style Chart EA
Here begins the main topic of this article.
I added a Nampin/Basket structure to the AI-style chart prediction EA.
The structure is simple.
Only the initial entry uses the AI-style chart判定 (判定: judgment).
If price moves against you, add in the same direction.
Increase lot sizes in steps via a lot-size sequence.
Recover with the entire basket.
In other words,
Use AI to narrow the entry.
Improve average entry price with Nampin.
Recover collectively with the basket.
This is the setup.
What’s important here is that I didn’t just add Nampin for its own sake.
The first entry is narrowed using AI-style chart判定.
That is, it does not enter indiscriminately at the first move.
Enter only when there is a certain directional sense based on similarity to past charts.
Then, if price moves against you, add in steps.
This combination worked quite well this time.
Strategy test results
The validation results this time were quite good.
The main results are as follows.