AI Similar Chart Prediction EA Development Notes 3 | Once I stopped matching the past, generic logic candidates began to emerge.
Hello, this is Tsumo.
This article isthe third note on AI Similar Chart Prediction EA Development.
Last time, I wrote about something rather bitter.
This is about an AI Similar Chart Prediction EA getting stuck in the swamp of past-data optimization.
Change the parameters.
The backtest numbers slightly improve.
Breaks under a different period.
Breaks under a different currency pair.
Change the parameters again.
Look at the numbers again.
This loop continues.
At first, development seemed to be progressing.
But I realized in the middle.
This isn’t about making a highly profitable EA for the future.
Aren’t you just looking for settings that fit past data?
So last time I paused it for now.
Continuing optimization in this direction would be dangerous.
Dousing the averaging in is also too early.
Rather than fitting to the past data of a single currency pair, seek a general logic that hardly breaks even across multiple currency pairs.
This time is about that rebuild.
To start with the conclusion, it is not finished yet.
Nor is it ready for actual operation judgments.
But the way of looking at it has shifted a little.
If you stop chasing past data fit, a candidate logic that is hard to break across multiple currency pairs becomes visible.
If you’re interested in MTP implementation, please take a look here as well.
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
Reflections from last time
The problem last time was simple.
But it was a deeply rooted issue.
Look at the EURUSD 15-minute chart results.
Adjust the settings.
The PF improves slightly.
Net profit improves.
Maximum drawdown decreases.
That part is fine.
The question is whether that improvement is genuine.
A particular currency pair.
A particular period.
A particular timeframe.
A particular setting.
There’s a possibility it only fits there.
That is not a practical operation logic.
In the past, it was strong.
In the future, weak.
Such an EA looks excellent on a backtest screen.
But in real operation, it suddenly becomes a different person.
Backtests are gentlemanly.
In real operation, they’re rough.
It’s the most troublesome type in EA development.
Stopping the search for “a winning setting” for now
What I did first this time was to stop looking for winning settings.
This is quiet but important.
Until last time, I inevitably looked at the numbers.
PF.
Net profit.
Win rate.
Maximum drawdown.
Number of trades.
Of course, these are important.
But focusing only on them tends to start looking for “settings with good numbers.”
And beyond that lies past-data fitting.
So this time I changed the perspective.
Aim to not maximize performance for a single currency pair.
Look for robustness across multiple currency pairs.
This is the policy I adopted.
I’m not aiming for the highest score.
First, look for resilience.
Not flashy, but solid foundational strength.
Starting over from here.
Evaluation criteria this time
The version evaluated this time is,AI Similar Chart Prediction EA V1.02e.
What differs greatly from last time is that optimization tailored to a single currency pair was stopped.
Only fit EURUSD.
Only fit USDJPY.
Only fit GBPUSD.
We will not do currency-pair-specific adjustments like this.
This time, we tried to operate with the same thinking across multiple currency pairs as much as possible.
The main conditions are as follows.
EA version: AI Similar Chart Prediction EA V1.02e
Timeframe: 15-minute chart
Lot size: 0.10 lots
Take profit width: 10 pips
Maximum holding time: 180 minutes
Time-based stop loss: after 60 minutes, if in drawdown, exit
Counter-signal settlement: enabled
Grid averaging: none
Currency-pair-specific optimization: none
The point of this time is not the precise numbers.
What’s important is,the fact that we looked at multiple currency pairs with a common approachtogether.
Rather than continuing to look for past-data-fitting settings like last time, this time we look at whether it resists breakdown.
Use of similarity judgment. But don’t over-detail it
This EA searches for shapes in the past chart that are close to the present.
Similarity judgments use mainly the following elements.
Price change.
Candlestick patterns.
Technical elements.
Time-of-day factors.
However, this article will not include exact parameter values.
The reason is simple.
If we write them too much, we’ll return to the question of “which settings are good.”
The theme this time is not finding the correct settings.
Not optimizing per currency pair, but seeing how far we can avoid breakdown with the same approach.
This is it.
We look at the design philosophy more than the settings.
Evaluation criteria this time
The evaluation criteria this time are quite simple.
We do not yet seek flashy profits.
First, we examine the following:
Whether it runs across multiple currency pairs
Whether PF does not break down drastically
Whether the number of trades is not too small
Whether maximum drawdown isn’t too heavy
Whether average losses are suppressed
Whether it does not rely on a single currency pair
Whether it can work without currency-pair-specific adjustments
In other words, we look at not the best peak performance, but
the minimum resilience.
This is the big difference from last time.