Choose Parameters on a Plateau, Not a Single Peak

Conclusion: Choose the Middle of the Range That Holds Up, Not the Single Best Result

Sort the optimization results by performance and adopt the top row. We consider this the shortest path to curve fitting.

What we look at is not the maximum value. It is whether performance holds up around the optimum. The range in which results stay smooth when you shift a parameter slightly is what we call a "plateau," and we adopt a value near its center. The numbers look less impressive than the single best point, and we are fine with that.

Why a Single Peak Is Dangerous

A Peak Contains "Lucky Hits"

Optimization searches hundreds to thousands of combinations for the one that best fits past data. The more trials you run, the more likely a combination that happened to catch a few large winning trades ends up near the top.

If performance drops sharply when you move to the neighboring value, we suspect that the result depends on a handful of specific trades rather than a real edge in the strategy.

The Market Won't Take the Same Shape as the Optimization Period

Future markets will differ from the past in volatility and in the cycle of price movement, little by little. This is similar to running with parameters that are slightly off.

Parameters whose neighborhood holds up have room to absorb some of that drift. A sharp peak has none.

Why We Use the Middle of the Range

If you choose the edge of a plateau, there is a cliff right next to you on one side. At the center, you can drift in either direction and still stay on the plateau.

What we are choosing is not how high the result is, but how far it is from collapse.

Verification Steps: Checking the Neighborhood in MT5

Checklist

1. Don't make the step too coarse: With a coarse step you can't tell a plateau from an isolated point. Once a promising range appears, rerun with a finer step around it only.
2. Pick two parameters and view them in 2D: MT5 can display optimization results as a surface. If dark cells spread across an area, it's a plateau; if only one cell is dark, it's a peak.
3. Don't judge by PF alone: Compare trade count and drawdown between neighbors as well. If PF holds but the trade count drops sharply, it may have turned into a different strategy.
4. Choose the middle of the range and record the gap from the maximum: Keeping a note of how much performance you gave up lets you verify the decision later.
5. Check on a period not used for optimization: A plateau is only a statement about the same data. Whether the shape holds outside that period has to be confirmed separately.

Be careful when using a genetic algorithm. Because the search concentrates on promising regions, it can finish without testing the neighborhood sufficiently. For checking the neighborhood, redoing it as a full search over a narrowed range is more reliable.

Basic MT5 operations are collected in EA & MT5 Know-How.

What You Can Do as a User

You can run the same check on EAs that are distributed to you. Just move the main parameters a little above and below their defaults, backtest, and see whether performance suddenly collapses.

For example, DONCHIAN ENGINE has measured backtest values of PF 1.5, 631 trades, and 9.4 years tested. The upgraded MEGAMAX DONCHIAN shows PF 1.55 and 629 trades, also from backtesting.

However, there is something we should state honestly here. These are results at the single adopted parameter point, and these numbers alone do not tell you whether the neighborhood is a plateau. A modest PF is not proof of a plateau. The work of moving parameters and checking cannot be replaced by published numbers.

Measured backtest values for each EA are listed in the measured ranking.

Limits and Cautions

• A plateau does not guarantee future profit. Even a wide plateau is only stability on the same past data. If the character of the market changes, the whole plateau can sink.
• Where the plateau ends is a subjective call. If you decide the acceptable drop in performance after the fact, in a convenient way, that is a form of curve fitting too. The criterion should be set before optimizing.
• With many parameters, you can't check everything. A 2D surface shows only two at a time. Breakdowns that involve three or more interacting parameters can be missed.
• With few trades, the shape of the plateau itself can't be trusted. A few trades can change the color of the surface.
• We ourselves can't claim this method is sufficient. In forward tests on demo accounts, several EAs are running at a loss. Even parameters chosen with the plateau in mind have not avoided losing phases.

Summary

• In optimization, look for a range that holds up around the optimum rather than the single best point
• Adopt the center of the plateau; prioritize distance from collapse over how high the result is
• Compare not only PF but also trade count and drawdown between neighbors
• For genetic algorithm results, recheck the neighborhood with a full search
• A plateau is stability on past data, not a promise of future performance
• A single published number can't tell you whether it sits on a plateau. Move the parameters yourself and check