Every backtest produces an optimal parameter set. The question most traders forget to ask is: how different would the result be with slightly different parameters? If bb_period=30 produces a Sharpe ratio of 15.0 and bb_period=32 produces a Sharpe of 14.5, the strategy is robust around its optimum. If bb_period=32 produces a Sharpe of 2.0, the optimum is a spike in a noisy landscape and the strategy is probably overfitted.
Parameter sensitivity is one of the most important and most overlooked aspects of strategy evaluation. It determines whether your backtest results will survive contact with live markets where parameters never match perfectly.
Why Perfect Parameters Do Not Exist
In backtesting, you have exact historical data and can find the parameter that produces the best result on that data. In live trading, three things change.
First, the market evolves. The volatility structure, correlation patterns, and liquidity profile shift over time. The parameter that was optimal for 2023 may not be optimal for 2026. If your strategy is sensitive to small parameter changes, these market shifts will degrade performance more than if the strategy is robust across a range of parameters.
Second, execution differs from backtests. Slippage, timing differences, and partial fills mean live trading operates at slightly different effective parameters than backtesting. Your backtest assumes entry at the exact candle close when the indicator threshold is crossed. Live trading enters a few seconds later at a slightly different price. If the strategy is sensitive to exact entry timing, these differences compound.
Third, you will inevitably choose the wrong parameter. Parameter optimization selects the best result from historical data, which includes both genuine signal and random noise. The noise component will not repeat in live trading. A robust strategy degrades gracefully when the noise does not repeat. A sensitive strategy collapses.
Measuring Sensitivity
The simplest sensitivity test is a parameter neighborhood analysis. Take your optimal parameter and test values within plus or minus 10 and 20 percent. Plot the Sharpe ratio across this range.
Our Bollinger Band strategy at bb_period=30 shows this profile on the original six altcoins: bb_period=25 produces Sharpe 8 to 10. bb_period=28 produces 10 to 14. bb_period=30 produces 12 to 17. bb_period=32 produces 11 to 16. bb_period=35 produces 10 to 14. bb_period=40 produces 14 to 22 (a second peak discovered in Phase 2). The performance landscape is smooth with broad plateaus rather than narrow spikes.
Compare this to what we observed with statistical strategies. The wavelet decomposition strategy at its optimal parameters (including DWT level, threshold, and reconstruction parameters) showed a sharp spike at the optimum with rapid degradation on either side. Changing the primary parameter by 15 percent cut the Sharpe ratio by more than half. This sensitivity was a warning sign that the strategy was fitting to specific frequency characteristics in the training data.
The Phase 2 Approach
Our parameter optimization uses a three-phase approach that explicitly addresses sensitivity. Phase 1 (initial sweep) tests a coarse grid of parameter values. Phase 2 tests a plus or minus 20 percent grid around the Phase 1 winner with finer resolution. Phase 3 tests a tight grid on validated symbols only.
Phase 2 specifically reveals parameter sensitivity because it explores the neighborhood of the optimum. If Phase 2 finds that nearby values produce similar results, the optimum is robust. If Phase 2 finds that nearby values produce dramatically different results, the optimum is fragile.
Our bb_period=48 discovery came from Phase 2. The Phase 1 winner for the new seven symbols was bb_period=30, but Phase 2 showed that bb_period=48 was a separate, higher peak. The landscape had two broad plateaus (around 30 and around 48) corresponding to two different liquidity groups. This was a genuine structural finding, not a narrow spike, because both plateaus were broad and robust to small parameter changes.
Sensitivity Across Symbols
Parameter sensitivity should be evaluated per-symbol and per-symbol-group. A parameter that is robust on SOL might be sensitive on PEPE because the two assets have different liquidity profiles and volatility structures.
Our two-group parameter scheme (bb_period=30 for liquid altcoins, bb_period=48 for thin altcoins) was developed specifically because the sensitivity analysis showed different optimal plateaus for each group. Within each group, the parameters are robust. Between groups, using the wrong parameters produces dramatically different results (including zero trades for bb_period=48 on the liquid group).
This means sensitivity analysis must be done at the symbol level, not the aggregate level. A strategy that shows low aggregate sensitivity might be hiding high sensitivity on specific symbols where the parameter landscape is uneven.
Sensitivity as a Signal Quality Indicator
Robust parameter sensitivity is an indicator of genuine signal. When a strategy works across a broad range of parameters, it is likely capturing a real market pattern that persists regardless of exact parameterization. The specific parameters tune the capture, but the underlying edge exists independently.
Fragile parameter sensitivity is an indicator of overfitting. When a strategy only works at one precise parameter setting, it is likely fitting to a specific feature of the historical data (a particular volatility regime, a specific news cycle, a temporary correlation pattern) that will not repeat.
Our deployed strategies all show broad, smooth parameter landscapes. Mean reversion works from bb_period 25 to 50 with varying optimality. Momentum works with rsi_period from 8 to 14. The optimal values are not arbitrary points in a noisy landscape. They are peaks in a smooth surface that reflects the underlying market structure.
Practical Sensitivity Testing
For any parameter optimization, add these steps to your workflow.
After finding the optimal parameter, test values at minus 20, minus 10, plus 10, and plus 20 percent. If any of these produce a Sharpe ratio less than half the optimal, the strategy is sensitive and the optimum may not survive live trading.
Plot the parameter landscape. Even a simple 2D plot of parameter value versus Sharpe ratio reveals whether you are on a plateau or a spike. Plateaus are robust. Spikes are fragile.
Test the optimal parameters on held-out time periods. If the parameters that were optimal on 2021 to 2023 data produce reasonable results on 2024 to 2026 data (even if not optimal), the strategy is robust. If they produce negative results on the held-out data, the optimization was fitting to the training period.
Use walk-forward analysis to verify that the optimizer consistently selects similar parameters across rolling windows. If the optimal bb_period is 30 in most windows but jumps to 55 in one window, investigate why. Instability in parameter selection across walk-forward windows is the walk-forward equivalent of parameter sensitivity in static analysis.
The goal is not to find perfect parameters. It is to find robust parameters that work well enough across a range of market conditions. The strategy with Sharpe 12 that is robust to plus or minus 20 percent parameter changes is more valuable than the strategy with Sharpe 18 that collapses when you change one parameter by 5 percent.