What is a walk-forward
A backtest picks a strategy's parameters and judges them on the same stretch of history. The result looks better than what comes next, because the parameters were fitted to that exact past.
A walk-forward separates the two. Parameters are chosen on one window (in-sample) and judged on the window right after it (out-of-sample), which they have never seen. Then both windows move forward and it repeats. Each repeat is a fold. The result you read is only the unseen windows, joined together.
In Luxis
Open Walk-Forward at the top of the strategy's chat and create a validation plan: In-sample months, Out-of-sample months, Years of history. The AI suggests 12 months in-sample and 3 out-of-sample. With fewer than 10 folds the verdict is statistically weak, and the panel says so.
For each fold:
- Run backtests on the fold's in-sample window, trying parameters. You or the AI can run as many as you like.
- Pick one run as the fold's winner.
- Go OOS: that run is tested on the fold's unseen window. You get one try. After it, the fold closes.
Luxis does not pick the parameters for you. The one try is what keeps the test honest: a second look at the unseen window would turn it into data you tuned on.
The code is frozen
A plan runs the strategy's code as it was at the plan's first backtest. If you change the strategy later, the panel tells you and asks whether to keep the plan's code or publish the change into the plan.
Reading the result
The Edge Report at the bottom of the Walk-Forward panel grows as folds close, and gives the verdict.
Rules that keep it honest
- Plan before the first backtest, not after a good one. A backtest over a fold's out-of-sample months before that fold was judged marks it "seen by exploratory BTs", and weakens its verdict.
- Never re-tune a fold that is closed.
- Never backtest the out-of-sample dates "just to see".
- Running several plans and keeping the best one is choosing, not validating.
After it
A walk-forward never finishes. You can go live from a plan: the live then runs the frozen code with the winner's parameters, and each next fold's out-of-sample is the live's real trading. After each live period you tune again on the months before it, and the live goes on. See From backtest to live.