Expanding Versus Rolling Training Windows

A strategy meant to trade for years can't just be trained once on historical data and left alone — markets change, so the model needs to be retrained periodically as new data arrives. The question is what training data to use each time: keep accumulating everything the model has ever seen, or only look at the most recent stretch. This choice — expanding versus rolling windows — trades off how much data the model gets against how current that data is, and it can meaningfully change how a walk-forward backtest performs.

Two ways to slide forward

An expanding window starts with an initial training period and grows it every time the model retrains, always including all history from the start date to the present. By the time the backtest reaches its last retraining date, the model has seen years, possibly decades, of data — maximizing sample size, which especially helps models with many parameters that need more data to estimate reliably. A rolling window keeps the training set a fixed length (say, always the trailing three years) and slides both ends forward together, dropping old data as new data arrives. This keeps the model trained only on relatively recent conditions, which helps if the market's underlying relationships genuinely change over time (a regime that mattered ten years ago may be actively misleading today), but it means training on less data at every step and discarding information that might still be useful.

Worked example: retraining a volatility-timing model

A model that adjusts position size based on recent realized volatility is retrained quarterly over a fifteen-year backtest. With an expanding window, the tenth retraining uses all 7.5 years of prior data, giving stable parameter estimates but potentially still weighting a very different low-rate, low-volatility regime from a decade earlier as heavily as recent data. With a rolling three-year window, the same tenth retraining uses only the most recent three years — if those three years happened to be unusually calm, the rolling model may underestimate volatility risk more than the expanding model would, because it has entirely forgotten the more turbulent years the expanding window still remembers. Backtesting both variants side by side and comparing their walk-forward Sharpe ratios and drawdowns is the standard way to see which sensitivity dominates for a given strategy.

What this means in practice

Neither approach is universally correct — expanding windows suit strategies where the underlying relationship is believed to be stable and more data reliably improves estimation; rolling windows suit strategies where market structure or regime genuinely shifts and stale data actively hurts. In practice teams often test both in a walk-forward backtest and compare, and some use a hybrid: expanding up to a cap, then rolling once the window reaches a chosen maximum length, balancing data volume against staleness.