Takuya Natsuno / Profil

**Can an EA Decide Not to Trade

When the Market Becomes Dangerous?**

Through long-term testing spanning nearly a decade,
one idea has gradually moved to the center of my EA design process:

The problem is not how to reduce drawdown itself,
but how an EA can recognize the conditions in which drawdown is most likely to emerge.

This post is not a final answer.
It is a record of design insights obtained during the development of an upcoming upgrade.

What Long-Term Testing Revealed About “Dangerous” Drawdowns

In a 10-year backtest, it was expected that drawdowns would be uneven across different years.
That part was not surprising.

What was unexpected was the existence of periods in which performance remained negative
for almost an entire year.

The most problematic environment turned out to be markets with
no clear directional bias, yet persistent and sharp price swings.

These were neither trend-following nor range-bound conditions.
Volatility was high, price moved aggressively in both directions,
but without producing meaningful structure.

In such environments, the previous model failed to react with sufficient precision.
In particular, the logic responsible for dynamically adjusting take-profit and stop-loss levels
was not adapting fast or accurately enough.

This made it clear that
the “mechanical brain” responsible for managing exits and risk itself needed to evolve.

Market Behavior: Accidental, Yet Inevitable

I do not see these environments as random anomalies.
They are accidental in appearance, yet inevitable in nature.

A market without surprise, expectation, or disappointment would require
a world in which humans are no longer human.
Such a market does not exist.

The real difficulty lies in the fact that
EA systems are structurally poor at detecting these conditions in advance.

An EA’s market interpretation is built upon accumulated past data.
As a result, it will always react after a regime shift has already begun.

This is not a flaw.
It is simply a structural condition of algorithmic trading systems.

Abandoning a Single Perspective

Accepting this limitation led me to reconsider designs
based on a single timeframe or a single analytical perspective.

When exploring alternatives, an image came to mind:
a creature with multiple heads and glowing eyes.

Each head observes the market independently,
yet none is allowed to dominate decisions on its own.
Tracing these perspectives back leads to a single core judgment.

Some perspectives observe market behavior across different time horizons.
Others focus on volatility, instantaneous price expansion,
or the effective width of price ranges.

By combining these independent “eyes,”
the core decision-making process becomes slower—but also deeper.

The goal is not to predict the market.
The goal is to avoid overlooking danger.

Changing Participation, Not Prediction

In the current development-stage model,
this design philosophy has produced a clear behavioral shift.

When comparing the strong JPY depreciation environment of 2024
with the highly unstable pandemic-driven market of 2020,
the number of executed trades differs by nearly a factor of four.

In particularly dangerous conditions like those seen in 2020,
the system intentionally suppressed trading activity,
successfully compressing drawdowns relative to earlier models.

This was not a change in how profits were pursued,
but a change in how survival was prioritized.

Why Drawdown Zero Is Not the Objective

Drawdowns cannot be eliminated entirely.
And attempting to suppress them too aggressively
can itself become destructive.

Reducing drawdown does not automatically lead to better performance.
The smoother the equity curve becomes,
the more likely overall return potential begins to decline.

If a system ultimately becomes less attractive
than a simple bank deposit,
then the research itself has failed.

For me,
the question of what it truly means to face risk
remains an ongoing challenge.

This project is still a work in progress.
But through continued design and verification,
I aim to better understand what makes an EA viable for long-term use.

Hat sich auf MQL5.community registriert