Walk-Forward Backtesting for Crypto Strategies: When Should You Lock In Parameters?

Trading Strategies

29 October 2026

Parameters should be locked in before entering out-of-sample testing. Once you start evaluating performance on out-of-sample data, any parameter adjustment based on that performance strips the data of its "out-of-sample" meaning. The value of walk-forward backtesting lies precisely in forcing you to make this decision at the boundary of each window: you can optimize repeatedly inside the training window, but once you cross the boundary into the testing window, parameters must be frozen, and test results are only recorded, never modified. If you go back and change parameters when the testing window performs poorly, it is no longer walk-forward backtesting—it is overfitting disguised as validation.

OKX Exchange
A leading global cryptocurrency platform,suitable for both beginners and experienced traders.
New user benefit: 20% off trading fees upon registration!！

When Does Locking In Happen?

Each fold in walk-forward backtesting has a clear training segment and a testing segment. Parameter selection happens only inside the training segment, and the testing segment is used only for evaluation. CoinPaprika describes walk-forward backtesting as fitting parameters on one window, applying those parameters unchanged to the next out-of-sample window, recording the results, then advancing both windows and repeating. VectorBT's example is more specific: the best parameter set selected in each training year achieved an average out-of-sample Sharpe ratio of only 0.24 in the following quarter, while the in-sample Sharpe ratio was 1.92. That gap is exactly the kind of information that only becomes visible after parameters are locked in.

So the precise meaning of "locking in" is this: the moment the training window ends, the parameters are fixed. No matter how good or bad the testing window results are, they cannot be used to adjust the same parameter set and then retest the same data. If you find that testing performance is poor, the correct action is to record the failure, move on to the next fold, and not go back to the previous training window to swap in a different parameter set.

Why You Cannot "Decide Parameters After All Backtests Are Done"

There is a common mistake: run all walk-forward windows, pick the parameter set with the best overall performance, and then claim it as the final set. This effectively includes the testing windows in the selection process. CoinPaprika's test data illustrates the problem clearly: among 88 RSI parameter configurations, 78 were profitable in-sample, but after subtracting benchmark returns, only 2 produced positive excess returns. If you pick the configuration with the prettiest overall curve from all 88, you will most likely pick the one that happened to catch the 2017–2021 bull market environment, not a rule that will remain effective in the future. The same data also shows that the unconditional 30-day return across those eight assets was 17.92% from 2017 to 2021, but only 2.40% from 2022 to 2026. The difference between time periods alone is far larger than the difference most parameters can make.

After Locking In, When Can You Unfreeze?

Parameter locking is not permanent. The forward progression of walk-forward backtesting is itself a form of controlled unfreezing: when the next fold begins and the training window moves forward, you can re-optimize parameters on the new training data. A strategy validation article from Gate lists this as one part of layered validation: in-sample backtesting is used to build the strategy and make initial parameter selection, out-of-sample testing checks generalization on data not used in optimization, and walk-forward testing repeats the examination across different time windows.

This means parameters can be updated as market conditions change, but the update must happen on new training data, not on testing data you have already seen. If you run the strategy live for a while and find that performance consistently deviates from backtest expectations, whether and when to adjust parameters depends on whether your strategy rules predefine "failure conditions." A strategy without predefined failure conditions cannot easily distinguish any parameter adjustment made after live losses from overfitting.

A Practical Verification Standard

Before locking in parameters, use two questions to check whether the parameter set deserves to enter out-of-sample testing.

First, are the parameters stable within the training window? If parameters jump wildly between adjacent training windows—for example, moving average periods changing from 10 and 200 to 20 and 60, then back to 10 and 200—that is itself a warning sign. CoinPaprika points out that instability of the best parameters jumping between windows is usually evidence against the rule. Stable parameters are not necessarily good parameters, but wildly jumping parameters almost certainly cannot be.

Second, have cost assumptions already been included? The same CoinPaprika dataset shows that a 1% round-trip cost is enough to turn the best +0.23 percentage point excess return into -0.77 percentage points. If your strategy has high turnover, cost sensitivity testing should be completed before locking in parameters, not after out-of-sample results come out. A strategy that is barely profitable under zero-cost assumptions will most likely fail under real costs.