Walk-forward backtest explained: how to test a trading model without cheating
A walk-forward backtest trains a model on history up to a date, forecasts the stretch after that date, then moves forward and repeats. Only the forecasts made on unseen data are scored, which avoids look-ahead bias, the most common way backtests look better than reality.
Why a plain backtest can mislead
If a model is scored on the same data it learned from, it can memorise the past and look brilliant. Look-ahead bias also creeps in when a rule, a threshold or a list of markets is chosen after seeing the results. Walk-forward testing removes the first problem by never scoring data the model has already seen.
How it works
Train on years 1 to N. Forecast year N+1 and record the results. Retrain including year N+1, forecast year N+2, and so on. Join all the unseen stretches together and you have one continuous out-of-sample record.
What it still cannot prove
It is a historical test. It does not include trading costs, slippage or your timing, markets can change, and the choice of which markets to include can still bias results. Accuracy is not profit. Treat it as evidence about forecast quality, not a promise of returns.
How ApolloAlpha uses it
ApolloAlpha's published record is walk-forward and out-of-sample: 72.7% of weekly calls were right across 163 markets, and 82.6% when the model was at least 70% confident. See the track record page for the method and what it does not show.
Frequently asked questions
What is look-ahead bias?
Using information in a test that would not have been available at the time of the forecast, for example training on future data or picking settings after seeing the results.
Does a good walk-forward result guarantee future profit?
No. It shows the forecasts were good on past unseen data. Costs, changing markets and your own decisions still matter.
Published 2026-11-07. See the track record and audit for the full method. Statistical model output, not financial advice. Past results do not guarantee future results.