Quant
Walk-forward validation
In an earlier post we saw the core defence against fooling yourself: tune on one slice of history, test on a slice you never touched. Walk-forward validation takes that honest idea and makes it far harder to game — by doing it over and over, marching through time.
The problem with a single split
Split your data once — tune on 2023–2024, test on 2025 — and you get exactly one out-of-sample verdict. But what if 2025 just happened to suit your strategy? One test period can flatter or punish you by luck. A single pass can't tell a robust edge from a fortunate one.
Slide the window forward
Walk-forward validation repeats the tune-then-test cycle in a rolling window. Fit on an early stretch, test on the period right after, then slide both forward and do it again — and again — until you've marched across all your history. Now you have a sequence of out-of-sample results, each on data the strategy hadn't seen when it was tuned.
Don't test once. Tune, test on the next slice, roll forward, and repeat — so the strategy has to keep earning its edge across many unseen periods, not just one lucky one.
Reading the results
- Consistency beats a single big number. An edge that shows up across most windows is far more trustworthy than one carried by a single spectacular stretch.
- Anchored vs rolling. You can keep the training start fixed and only grow the window (anchored), or slide a fixed-length window along (rolling). Rolling adapts to changing regimes; anchored uses more history.
- It's stricter, and that's the point. Walk-forward will happily kill strategies that a one-shot backtest called winners. Better to learn that on your laptop than with real money.
It's more work than a single split, and it should be. The whole game is separating a durable edge from a comfortable story — and few tests apply that pressure as honestly as walking the strategy forward through time.