Parameter plateaus beat parameter peaks

[Fairman]

Parameter plateaus beat parameter peaks

An optimizer produces a table: stop distance, target, filter threshold, and profit for each combination. The temptation is to pick the best row.

The best row is often a spike. Move one parameter slightly, and the profit collapses. That signals a fit to noise.

A more robust choice sits on a plateau: a region where nearby values also perform well. If a 4 pip stop earns 100, and 3 and 5 pips earn 90 and 95, the area is stable. The exact value matters less, and small changes in market behavior won't break it.

For scalping, where costs and conditions shift week to week, robustness matters more than peak performance. A parameter set that earns 80 percent of the maximum across a wide range will survive drift, while the peak setting often won't.

Choosing from the center of a plateau also gives a built-in margin against cost increases, such as a wider spread or extra slippage.

When was an optimized setting last chosen from the middle of a plateau instead of the best single cell?

[Shadow Trader]

Optimizing an indicator period from 5 to 20 and picking the best value is standard practice. The winner might be 7. The neighbors, 6 and 8, show much worse results. That's a red flag.

If a strategy is robust, nearby parameters give similar results. Performance changes gradually, forming a smooth surface. If one value works and its neighbors fail, the system is fitted to noise. The 7 found the quirks in the historical data, and they won't recur.

Scalping strategies are especially vulnerable. The number of trades is large, which creates an illusion of statistical power, but the underlying edge is thin, and the optimizer finds noise that happens to align.

A sensible habit is to examine the whole parameter landscape. Plot results against parameter values. Look for plateaus, not peaks. Choose a value in the middle of a good region, not the highest spike.

Another check: add a small amount of random noise to the data or the parameters and see whether the results survive.

How wide is the plateau around the chosen parameters in the strategy currently in use?