Walk-forward analysis

Why a single split is not enough

A 70/30 in-sample / out-of-sample split is a real test, but it judges the strategy on exactly one stretch of future. Pass it and you have one piece of evidence; fail it and you might just have caught one bad regime. Walk-forward turns that single test into many: it slides the whole optimize-then-test process forward through history, so the strategy is validated on window after window of data it never optimized on.

How walk-forward works

Split the history into a sequence of folds. In each fold you optimize on a training window, then test the resulting settings on the test window that immediately follows, then move both windows forward and repeat. You stitch the test segments together into one continuous out-of-sample equity curve. That stitched curve, built entirely from data the optimizer never saw at the time, is the honest estimate of how the strategy would have performed.

window:
  train_months: 18      # optimize on this much history
  test_months: 6        # then test on the next slice, unseen
  roll_months: 6        # slide both windows forward by this much
folds:
  - { train: 2019-01..2020-06, test: 2020-07..2020-12 }
  - { train: 2019-07..2020-12, test: 2021-01..2021-06 }
  - { train: 2020-01..2021-06, test: 2021-07..2021-12 }
stitch: concatenate test segments into one OOS curve

Anchored versus rolling windows

There are two flavours. A rolling window keeps the training length fixed and slides it forward, so the strategy always tunes on the most recent N months and adapts to changing regimes. An anchored window keeps the training start fixed and lets it grow, so the strategy tunes on all history to date. Rolling adapts faster and is better for markets that change character; anchored is more stable and uses more data. Try both and see which the strategy prefers.

• Rolling: fixed-length training window, slides forward, adapts to regime change.
• Anchored: fixed start, growing training window, more data and more stability.

Reading walk-forward efficiency

Compare the stitched out-of-sample performance to the in-sample performance the optimizer found in each fold. If out-of-sample is a reasonable fraction of in-sample, the strategy generalises and you can believe it. If out-of-sample collapses to near zero while in-sample looked great, the optimizer is fitting noise every fold and the edge is imaginary. A few weak folds among many strong ones is normal; consistent out-of-sample failure is the verdict to act on.