Parameter Stability

How to read the parameter trajectory chart and stability table -- CV explained, stability tiers, what stable and unstable parameters mean, and which values to use going forward.

What is Parameter Stability?

When you run a walk-forward analysis, the optimizer finds the best parameter values independently in each training window. Parameter stability asks a simple question: did it keep picking similar values, or did it pick wildly different ones each time?

If the optimizer picks an RSI period of 28, then 30, then 29 across three windows, that parameter is stable -- the optimizer is converging on the same region of the search space regardless of which slice of data it trains on. If it picks 15, then 42, then 22, the parameter is unstable -- the optimizer is latching onto different patterns in different time periods.

Stability matters because a parameter that changes dramatically from window to window is unlikely to work reliably going forward. If there is no consistent optimal value, the optimizer is probably fitting noise rather than capturing a real market property.

Same parameter (RSI period) across 6 walk-forward windows. The stable case clusters tightly; the unstable case scatters across the range.

Coefficient of Variation (CV)

Parameter stability is measured using the coefficient of variation (CV) -- the standard deviation of the optimal values divided by their mean, expressed as a percentage.

Coefficient of Variation

CV = (standard deviation of optimal values / mean of optimal values) x 100

A low CV means the values cluster tightly around the average. A high CV means they are spread out. Because CV is relative to the mean, it works across parameters with different scales -- you can compare the stability of an RSI period (range 5-50) with a Bollinger Band multiplier (range 1.0-3.0).

Worked Example

Suppose the optimizer selected RSI period values of 28, 30, 29, 31, 28, and 30 across six windows. The mean is (28 + 30 + 29 + 31 + 28 + 30) / 6 = 29.33. The standard deviation is 1.21. The CV is 1.21 / 29.33 x 100 = 4.1%. That is very stable.

Now suppose the values were 15, 42, 22, 38, 12, and 35. The mean is 27.33 and the standard deviation is 12.68. The CV is 12.68 / 27.33 x 100 = 46.4%. That is variable -- the optimizer is finding a very different optimal value each time.

Stability Tiers

Simula assigns each parameter a tier based on its CV. These are guidelines, not hard rules -- what matters is the pattern, not whether a parameter lands at 9% or 11%.

CV below 10% -- Very Stable

The optimizer converges on similar values across windows. This is strong evidence that the parameter captures a consistent market property.

CV 10-25% -- Stable

The values cluster in a recognizable range. The parameter is likely meaningful and captures a real pattern.

CV 25-50% -- Moderate

Noticeable variation. The parameter has some consistency but shifts across market conditions. Worth investigating whether narrowing the sweep range improves stability.

CV above 50% -- Unstable

The optimizer is choosing very different values each time. This parameter is likely fitting noise rather than signal. Consider removing it from the optimization or fixing it to a sensible default.

Stable parameters are the strongest evidence of a real edge

When the optimizer independently arrives at similar values across different time periods and market conditions, it is finding the same pattern each time. That convergence is one of the strongest signals that a strategy has a real, repeatable edge -- arguably stronger than the WFE number itself.

Reading the Parameter Trajectory Chart

The parameter trajectory chart plots each swept parameter's optimal value across windows as a line, normalized to a percentage of its full sweep range. A parameter swept from 10 to 50 that the optimizer set to 30 would appear at 50% of the chart height.

The normalization uses the sweep range you defined (the start and stop values), not just the values the optimizer happened to pick. This means a stable parameter that only uses a narrow band of its range appears as a tight horizontal line, while an unstable one zigzags across the full chart height.

What to Look For

Flat or near-flat line. The optimizer picks similar values every time. The parameter is stable and likely captures a consistent market property.

Steady drift in one direction. The optimal value is gradually shifting over time. This suggests the market regime is changing -- the parameter is adapting to evolving conditions. This is not necessarily bad, but it means the most recent window's value is probably more relevant than the mean.

Random bouncing. The values jump around with no pattern. The optimizer is finding different optima in each period and the parameter is not stable. The optimization surface for this parameter is likely noisy.

Reading the Stability Table

The parameter stability table shows five columns for each swept parameter:

Last Window

The optimal value from the most recent training window. This is the value trained on the freshest data and is generally the best choice for forward use -- it reflects current market conditions. This is the approach recommended by Robert Pardo, who formalized walk-forward analysis.

Mean

The average of the optimal values across all windows. When parameters are stable (low CV), the mean is a good consensus estimate that smooths out window-to-window noise. When parameters are unstable, the mean is less meaningful because the values it averages are too different from each other.

Standard Deviation

How much the optimal values vary around the mean, in the parameter's own units. A standard deviation of 1.2 on an RSI period with a mean of 29 is tight. A standard deviation of 12.7 on the same mean is wide. The CV puts this in relative terms.

The coefficient of variation -- standard deviation divided by the mean, expressed as a percentage. This is the number the stability tier is based on.

Stability Tier

A qualitative label (Stable through Unstable) based on the CV thresholds described above.

Choosing Which Parameters to Apply

The stability table includes two apply buttons: one for the last window's parameters and one for the mean. Which should you use?

When parameters are stable (low CV): The last window and the mean will be very similar -- either is a good choice. The mean smooths out minor noise; the last window reflects the most recent data.

When parameters are variable (high CV): The last window's parameters are the safer bet. They reflect the most recent market behavior. The mean is an average of very different values and may not represent any real optimal point.

When the trajectory shows a drift: The last window is clearly better. The mean would pull you back toward outdated values from earlier periods that the market has moved away from.

Can I re-optimize over the full date range instead?

Yes. If WFE is comfortably above 50%, you can run a full optimization over the entire date range with greater confidence. The walk-forward result gives you permission to trust the optimization result more than you otherwise would. This is the most aggressive approach and should only be used when WFE clearly passes.

What to Do With Unstable Parameters

An unstable parameter is not just a statistical curiosity -- it is a signal that something about the strategy needs attention. Here are concrete actions you can take:

Narrow the sweep range. If the optimizer is bouncing between 10 and 45, but domain knowledge says values below 20 are unreasonable, constrain the range to 20-45 and re-run. A smaller range gives the optimizer less room to overfit.

Fix the parameter to a sensible default. If the parameter does not have a stable optimum, it may not be contributing meaningful information. Fix it to a conventional value (e.g., RSI period of 14) and remove it from the optimization entirely. Fewer swept parameters means fewer degrees of freedom and less risk of overfitting.

Question the indicator itself. If the parameter for an indicator is consistently unstable, the indicator may not be contributing a useful signal for this asset and timeframe. Consider replacing it with a different indicator or simplifying the strategy.

Do not ignore unstable parameters

It is tempting to focus on the WFE score and skip the stability analysis. But a strategy can pass the WFE threshold while relying on parameters that change every window. That means the strategy's edge is fragile -- it works, but only if you re-optimize frequently and happen to pick the right values. Stable parameters give you confidence that the edge will persist without constant re-tuning.

Where to Go Next

Reading Walk-Forward Results -- overview of the full results page, including the other components beyond parameter stability.

Walk-Forward Efficiency -- the WFE metric and what it tells you about overall strategy robustness.

Grid Search & Overfitting -- how parameter count and sweep range affect overfitting risk.

Overfitting & Curve Fitting -- the broader problem that parameter stability helps detect.