Backtesting
A single split isa number you cantune until youlike it
Optimise on the whole history and the result is a description of that history. Walk-forward fits on one window, tests on the next, and rolls — so every result you read was produced by parameters chosen without seeing it.
The mechanic, and what it protects against
Fit, test, roll
Choose parameters on an in-sample window. Evaluate them on the out-of-sample window immediately after it, which the fitting never saw. Move both windows forward and repeat. What you report is the concatenation of the out-of-sample pieces — a track record assembled from decisions that were made in advance each time.
It costs you the best-looking number, and that is the point
Walk-forward results are almost always worse than a single optimised split, and people quietly abandon it for exactly that reason. The worse number is the honest one: it is what the strategy would have produced if you had been running it rather than describing it afterwards.
It exposes parameter instability
If the best parameters change wildly from window to window, you have not found a rule — you have found a curve fit that re-fits. That instability is invisible in a single split and obvious across a walk, and it is the most useful thing the method tells you.
What it cannot fix
Walk-forward controls for one form of overfitting. It does not fix look-ahead in your features, fees you did not model, slippage at your size, or a market regime that simply is not in your data. Those are separate problems and each has its own page here, because bundling them into "we backtest properly" is how a platform sounds thorough and is not.
Questions about walk-forward
How long should the windows be?
Long enough that the in-sample window contains more than one regime, and short enough that you get several out-of-sample pieces. With the two years of history available here, that constrains the number of folds in a way worth knowing before you design the test.
Should the windows overlap?
The in-sample windows can roll with overlap; the out-of-sample pieces must not. Overlapping the pieces you report double-counts the same days and makes the track record look longer than it is.
What do I do with the result?
Compare it against doing nothing and against the cost of trading it. A walk-forward equity curve that beats neither is a strategy that has been tested honestly and failed, which is a good outcome for a test to produce.
Test something you already believe
Free paper accounts run the backtester on real history.