Backtest Lab — Walk-forward
Roll train/test windows forward through history to get out-of-sample metrics per window and detect overfitting.
Walk-forward analysis splits your window into rolling train and test (out-of-sample) segments, steps them forward through history, and reports metrics per test window. It’s the Lab’s overfitting check: parameters that only look good in-sample tend to fall apart on the test segments.
How the windows are built
The Lab constructs a sequence of rolling windows from your base config’s
startDate/endDate:
- Each window has a train span followed by a test span.
- Windows roll forward so successive test segments cover later, unseen periods.
- Date math is pure UTC calendar arithmetic — month addition clamps
day-of-month overflow (Jan 31 + 1 month → Feb 28/29, never Mar 2/3), and every
window’s start is derived from the original
startDateso rounding never compounds into drift.
Reading the results
You get an out-of-sample summary across windows plus per-window metrics. What to look for:
- Consistency — are the test-window metrics stable, or all over the place?
- In-sample vs. out-of-sample gap — a big drop from train to test is the classic overfitting signature.
- Worst window — a strategy is only as dependable as its bad periods.
Workflow
Walk-forward is the third step of a disciplined Lab loop:
- Sweep to find promising parameters.
- Walk-forward to confirm they survive out-of-sample.
- Compare the survivor against baselines.
Related
- Lab overview
- Learn — the concept of overfitting, taught interactively