Parameter Plateaus: Why a Stable Region Beats the Best Single Setting

[FXS Prop Desk]

Optimisation often finds a "best" parameter set. If nearby values perform much worse, the result is likely luck.

What to look for:

• A plateau: a range of parameter values with similar results.
• Smooth changes in results across parameters.
• Choosing a value in the middle of the plateau, not at the peak.

Methods:

• Heatmaps of results across two parameters.
• Sensitivity tables.

Strategies with few parameters and wide plateaus tend to hold up better in live trading.

A useful habit is to look at the worst result inside the plateau, not the best. If even the weakest neighbouring setting is acceptable, the strategy is less likely to disappoint when conditions change slightly.

Walk-forward results can confirm a plateau: if the chosen value stays inside the stable region in most windows, the setting is far more trustworthy than one that only works in a single period.

Do you check parameter stability when optimising?

[Shadow Trader]

FXS Prop Desk, I check stability, but I learned the hard way that a plateau under one cost assumption can vanish under another. My breakout EA had a flat region for the lookback between 14 and 22 bars at 0.6 pips of spread, and at 1.2 pips half of that region turned negative.

So my heatmaps now come in pairs, one at normal costs and one at roughly double. If the plateau survives the expensive version I take the middle of what is left, not the middle of the original.

I also agree on judging by the worst neighbour. My rule is that every cell within two steps of the chosen value must keep a profit factor above 1.15, and if one cell drops below that I treat the whole area as suspect.

One thing I would add is the number of trades per cell. A cell with 40 trades can look like a peak purely from noise, so I grey out anything under 150 trades before reading the map.

Do you optimise more than two parameters at a time? Past two I find the plateau hard to see, so I start fixing one parameter by logic instead.