Curve fitting of EA (overfitting) countermeasures

Frequently asked questionsDo you have test results from around the year 2000?We receive this question, but by optimizing, any currency can achieve excellent test results.

With a single EA, even if it covers all five currencies, creating an EA that yields a rising, explosive profit graph from test results dating back to around 2000 takes some time, but it is possible using our own methods.

This can be considered a specialized skill, but even across all 10 currencies at once, it is possible to create an EA with a rising curve.

However, the problem begins here.

Even an EA with an extremely optimized median, in real operation, may experience subtle discrepancies, andit may be difficult to produce consistent results.

If the historical test period is longer, reliability for the future is higher, certainly, but from experience, test results are not always reliable and should not be overtrusted.

An EA optimized for a fixed period in the past is calledcurve fitting, and

・It works only under certain market conditions and becomes useless when the period changes.

・Test results can be unusually good, but they are completely useless in real trading.

・For example, if the tested currency's period was a rising-buyer market, in real markets a selling scenario might be obvious, but tests would repeatedly classify it as a buy.

If the test period is biased toward either buy or sell,even with excellent test results, the EA that is produced will be biased.

In other words, because the test period must be specified,it is almost impossible to test optimization with a perfectly balanced buy and sell period.

For nearly five years, I have pursued “finding a universal median across all cases and thoroughly combining various indicators to optimize.”

While it is far better than making trades with no basis,

even if you fit formulas to past markets, they remain incomplete and unreliable... the past is the past...

This is the conclusion I arrived at.

However, on the other hand,

to win in future markets, a fixed rule-based logic is necessary

and it mustat least work in the past as well.

In other words,to prove whether the logic works in the future, rigorous historical testing is still required..

What should we do to solve this delicate balance? How can we create an EA that keeps winning?

After five years of trial and error, finally,

we achieved past verification andan EA that also works in real markets

has been completed.

Curve-fitting (over-optimization) measures are

in the pastwe have not performed excessive optimization (fitting).We have not overfitted.We deliberately do not make test results overly favorable.

Not temporarily, but to produce consistent results over the long term,we have

developed an EA that does not fixate on the past and can fundamentally judge the market.

And five years of past verification was not wasted.

Because I tried and combined various indicators, I learned what to combine from entry to take-profit to be effective.

It is fully possible to program a fixed method of entry, (logic) and turn it into an EA, and

the EA is superior in every aspect to discretionary trading (self-entry) in my view.