Rolling Window Forecasting: Window Size as a Bias-Variance Trade

Every rolling window is a bet that the recent past resembles the near future and the distant past does not. Widening the window buys you noise cancellation and pays for it in staleness, and both halves of that trade have closed forms you can compute before you fit anything.

Two different things are called rolling

Before the arithmetic, a disambiguation that causes real confusion. “Rolling window” names two separate ideas and they answer different questions.

• A rolling estimation window is a modelling choice: fit on the last m observations only, discard everything older. The alternative is an expanding window, which keeps all history. This is the one with a bias-variance trade in it, and it is what this page is about.
• Rolling-origin evaluation is a validation procedure: step an origin forward through the series, forecast h steps from each origin, and average the errors. Leonard Tashman’s 2000 review in the International Journal of Forecasting is the standard reference for why this, and not a random train/test split, is the correct way to score a time series model.

You can use one without the other. The rest of this page holds the evaluation fixed and asks only what m costs you.

The variance term shrinks as 1/m

Take the simplest rolling forecaster there is: the forecast for the next period is the mean of the last m observations. Suppose the series is a level plus independent noise with standard deviation σ. The mean of m independent draws has variance σ² / m. That is the entire variance term, and it falls hyperbolically: going from 3 points to 12 quarters the variance, but going from 12 to 48 only quarters it again. The returns to a wider window are steeply diminishing from the start.

The independence assumption is doing work here and it is worth knowing where it breaks. If consecutive residuals are positively correlated — which they usually are, because a model that has not captured a slow driver leaves that driver in the residual — the effective sample size is smaller than m and the variance falls more slowly than 1/m. The same correction appears in the aggregation arithmetic on choosing a forecast granularity, for the same reason.

The bias term grows linearly in m

Now let the level drift. Say it moves by a constant d per period. The mean of the last m observations estimates the level as it was at the centre of the window, which sits (m − 1) / 2 periods in the past. Forecasting one step ahead adds another period of drift. So the one-step bias is:

bias(m) = d · ( 1 + (m − 1)/2 )
        = d · (m + 1) / 2

Linear in m, and squared in the error. That is the whole shape of the problem: variance falls like 1/m, squared bias rises like m², so mean squared error is

MSE(m) = [ d·(m + 1)/2 ]²  +  σ² / m
         \_____ bias² _____/    \_ var _/

a convex function with an interior minimum. There is a best window width and it is neither the shortest nor the longest available.

Three widths on one series

Put numbers in. These two are the assumptions and everything below is arithmetic over them: a per-period drift of d = 0.5 units and a noise standard deviation of σ = 4 units — a series where the week-to-week noise is eight times the weekly trend, which is the ordinary situation in retail demand.

m = 3    bias = 0.5·(3+1)/2  = 1.00   bias² =  1.00   var = 16/3  = 5.33   MSE =  6.33
m = 6    bias = 0.5·(6+1)/2  = 1.75   bias² =  3.06   var = 16/6  = 2.67   MSE =  5.73
m = 12   bias = 0.5·(12+1)/2 = 3.25   bias² = 10.56   var = 16/12 = 1.33   MSE = 11.90

Three observations from those nine numbers. The middle width wins, and it wins narrowly — 5.73 against 6.33 is a difference of about 5% in root mean squared error, which is not the kind of gap that survives being measured on a hundred noisy series. The widest width loses badly, and it loses to bias: at m = 12 the squared bias is eight times the variance, so the estimator is not noisy, it is confidently out of date. And the composition flips completely across the range: at m = 3 variance is five times the squared bias, at m = 12 the ratio has inverted.

Working the minimum out properly, the same formula puts the optimum near m = 5 for these inputs, with an MSE of about 5.45. Which matters less than the shape: the curve is flat around its minimum and steep on the long side. Erring short costs you a little; erring long costs you a lot. That asymmetry, not the exact optimum, is the transferable result.

Choosing a width without a grid search

You can differentiate the expression above and solve for m, but for real series the drift is not constant and σ is estimated, so a closed-form optimum is false precision. Four constraints do more work than the algebra:

• The window must cover a whole number of seasonal periods. A 20-day window on daily data with a weekly cycle contains six Mondays and five Sundays, so the mean itself carries a day-of-week bias that has nothing to do with drift. This is the single most common way a rolling window goes wrong, and it does not show up in the decomposition above because that model had no seasonality in it.
• The window must not span a structural break. A break makes the drift assumption false in a way no width fixes: the pre-break observations are not stale versions of the current level, they are samples from a different process. Detect it first — see detecting structural breaks — and truncate at the break rather than tuning around it.
• The window must be long enough for the parameters. A rolling mean has one parameter and survives m = 5. A seasonal ARIMA with a dozen has no business seeing fewer than a few hundred points, and a rolling window that starves it produces coefficients that swing wildly from origin to origin — which reads as a volatile forecast and is actually an under-identified fit.
• Widen the window as the horizon grows. The bias term above was derived for one step. At horizon h it becomes d · (h + (m − 1)/2), so the drift-induced error is already large before m contributes anything, and the marginal cost of a wider window falls. Long-horizon forecasts want more history than short ones, which is the opposite of most people’s instinct.

If you do search over widths, search with a rolling origin and average across origins, and treat differences under about 5% as noise. And re-run the search on a schedule rather than once: the optimum depends on the ratio of drift to noise, and that ratio moves. A width chosen during a stable year is the wrong width the year demand becomes volatile, and nothing in the pipeline will tell you.