Parameter stability (plateau)

Parameter stability is the property of a trading rule whose results stay close to the chosen configuration when each parameter is moved one step on the grid. A plateau of similar neighbours suggests the rule works around its parameters; an isolated peak suggests the chosen values were fitted to one sample.

Senzoukria · Glossary · Updated September 2026

What the plateau test looks at

A parameter sweep produces one result per configuration: a lookback, a threshold, a stop distance, and so on. The optimiser keeps the configuration with the highest objective. Parameter stability asks a different question about the same grid: what happened to the immediate neighbours of that winner, one step away on one axis at a time?

Three shapes are possible. A plateau is a region where neighbours perform nearly as well as the chosen point. A slope is a region where results degrade steadily in one direction, which usually means the sweep stopped at the edge of its range. A peak is a chosen point whose neighbours are markedly worse: the rule only worked on those exact values.

  • The comparison uses the same objective as the sweep (expectancy, SQN or Sharpe per trade), so a plateau on one objective is not automatically a plateau on another.
  • Only adjacent grid steps are compared. A coarse grid can hide a narrow peak between two steps; a fine grid can show a plateau that is really noise around a flat surface.
  • Stability is a property of the surface, not a certification of the strategy. A flat surface of losing configurations is still a losing rule.

Why a peak is treated as an artefact

Markets do not respect the exact value of a lookback. Volatility, session structure and contract rolls shift the conditions that made one parameter value optimal. If the result collapses when a threshold moves by one step, the market will produce that step on its own. A plateau tolerates that movement; a peak does not. This is why the plateau shape is read as evidence that the rule captures a behaviour rather than a coincidence of the sample.

In Senzoukria

The backtest panel of the desktop shows a chart titled Parameter stability after a sweep or a walk-forward run. It plots the chosen configuration against its immediate neighbours on the grid, labels the shape as plateau, slope or peak, and states the neighbours' median as a percentage of the chosen result. The help text of that chart states the rule directly: a plateau, where neighbours do nearly as well, is what a real edge looks like; a peak is an artefact. When no sweep was run, the chart reports that there is nothing to compare the chosen configuration with.

The panel also caps the parameter grid and warns that every extra combination is one more draw in the same data. Stability, the deflated Sharpe and the probability of backtest overfitting are complementary readings of the same sweep.

Common mistakes

  • Reading the chosen point alone and ignoring the neighbours that the sweep already computed.
  • Widening the grid until a plateau appears, which changes the number of trials and inflates selection bias.
  • Treating a plateau on in-sample data as an out-of-sample result. The walk-forward folds answer that separate question.
  • Sweeping several parameters at once and reading a plateau on one axis while another axis sits on a peak.

This page in other languages

Frequently asked questions

How many neighbours are compared in a parameter stability check?
The check compares the chosen configuration with the configurations one grid step away on each parameter axis. With two swept parameters that is up to four neighbours; with one parameter it is two. The comparison depends on the step size you set for the sweep, so the same rule can look like a plateau on a fine grid and a slope on a coarse one.
Does a plateau prove the strategy will keep working?
No. A plateau says the result was not sensitive to the exact parameter values inside the tested sample. It says nothing about regime changes, costs that were not modeled or data the test never saw. It is one robustness reading to combine with walk-forward results, the deflated Sharpe and a permutation test.
What should I do when the chart shows a peak?
Treat the chosen configuration as fitted to the sample rather than as a rule. Look at the objective values of the neighbours, reduce the number of swept parameters, and rerun on a different period. If the peak persists across periods, the rule probably depends on a detail of the data rather than on a market behaviour.

Keep reading

Strategy researchTest a trading strategy, from observation to challenge rulesDefine an observation, question it with the copilot, write a rule and explore costs, drawdown and fictional challenge limits in an interactive example.Quant · Start hereQuant trading for beginners: from idea to backtestDefine a testable rule, choose suitable data and evaluate later periods after costs.Order flow · HubOrder flow trading: the complete guideThe hub. What order flow is, what it shows — aggression, delta, imbalance, absorption, liquidity — and the tools that read it, each linked to a full guide.