Forecasting method: moving averages

Forecasting · ~6 minute read

The simplest forecast that earns its keep

The moving average is the first forecasting method most planners meet, and the one most experienced planners still use more than they admit to. It is unfashionable next to exponential smoothing, ARIMA, and machine-learning approaches, but it has properties — transparency, robustness, ease of explanation — that make it the right tool for several genuine problems. This article walks through the three variants worth knowing (simple, weighted, centred), the cases where moving averages earn their keep, the cases where they consistently fail, and the practical settings that make the difference between a useful baseline and a slow-moving guess.

What a moving average is

A simple moving average forecasts the next period as the mean of the last n periods. A four-week moving average of weekly volume averages the last four weeks; a seven-day moving average of daily volume averages the last seven days. The longer the window, the smoother the forecast but the slower it adapts to change. A planner choosing a window length is choosing how much weight to put on the most recent data versus how much to put on stability.

A weighted moving average gives more recent periods more weight. A three-period weighted average with weights 0.5, 0.3, and 0.2 puts half the forecast on the most recent observation. This is a compromise between the simplicity of the simple moving average and the adaptive behaviour of exponential smoothing — in fact, an exponential smoothing model is, mathematically, a weighted moving average with the weights declining exponentially into the past.

A centred moving average is used differently — not as a forecast but as a way to extract the underlying trend from noisy historical data. It averages each observation against the observations either side of it (so a five-week centred moving average sits in the middle of a five-week window). It is useful for analysing what has happened, not for predicting what will happen, because by definition it requires future data the planner does not have.

Where moving averages earn their keep

Three situations consistently reward a moving-average approach.

Long-range capacity planning. When the planner needs a defensible volume estimate for the next financial year, a moving average of the last several years, with some adjustment for known structural change, is harder to beat than it looks. The error bars on any one-year-ahead forecast are large enough that exotic methods rarely outperform a well-chosen moving average. The simplicity also makes the assumption visible to finance and senior management, which protects the planner when the eventual reality differs from the forecast.

Stable, non-seasonal queues. Some queues simply do not have strong seasonality or trend. Back-office work in a steady-state operation is the clearest example. A four- to eight-period moving average is often as accurate as anything more complex, with none of the overhead.

As a baseline. Every forecasting team needs a naive baseline against which to measure their fancier model. A naive seasonal moving average (this Tuesday = average of the last four Tuesdays) is the standard benchmark. If the operation’s sophisticated model does not beat this naive baseline by a meaningful margin, the sophistication is not adding value and probably needs reconsidering. The simple volume forecaster on this site reports both the Holt-Winters fit and the naive seasonal baseline so the comparison is visible.

Where moving averages fall short

The method has well-known weaknesses. The first is lag: a moving average always trails the underlying trend. If volume is genuinely rising at five percent a month, a four-period moving average will systematically under-forecast. The longer the window, the worse the lag. The fix is either to shorten the window (at the cost of more noise) or to add a separate trend component (which is what exponential smoothing and Holt-Winters do).

The second is seasonality blindness. A simple moving average does not see weekly or seasonal patterns. A seven-day moving average of daily volume gives a single number that washes out the busy Monday and the quiet Sunday. The fix is to use a seasonal moving average — predict this Tuesday using the average of recent Tuesdays, not the average of the last seven days — but this only works if you have enough seasonal history.

The third is insensitivity to known events. A moving average has no way of knowing that next Monday is a bank holiday or that a marketing campaign starts on Thursday. The planner has to layer overrides on top, which is fine, but means the moving average is only half the answer.

Practical settings that work

A few habits separate well-used moving averages from carelessly-used ones. Choose the window deliberately. Four periods is the operational default; longer windows are for stable queues; shorter windows are for queues with rapid change. Document the choice. Use seasonal moving averages where seasonality matters. A four-week moving average of Tuesdays out-performs a four-week moving average of all days for almost every voice queue. Smooth one-off events before averaging. A volume spike from an unrelated incident will distort the average for as many periods as the window length. Either exclude the spike or replace it with an estimate of what would have happened otherwise. Refresh on a fixed cadence. Roll the window forward weekly, not opportunistically when the planner remembers, so the forecast is always built from the same recipe.

Combining with overrides

A moving average is rarely the complete forecast on its own. It is the baseline; the planner adds overrides for known events, business changes, and emerging trends the model has not yet absorbed. The discipline is to keep the baseline pure (it is what it is, no fiddling) and to log the overrides separately, so the operation can later review whether the overrides were right. This is the same pattern used with more complex models, and the team that builds the habit on moving averages carries it forward easily as their methods grow.

Common mistakes

Four patterns recur. Choosing the window length without thinking, usually using whatever the previous planner used. Forgetting seasonality and producing a single moving average across days of the week. Mixing observed and forecasted data into the same window so the forecast becomes self-referential. And, most commonly, treating the moving average as the answer rather than as a baseline that needs to be augmented with business intelligence and event overrides.

Conclusion

Moving averages are not glamorous, but they are robust, transparent, and surprisingly effective for several common planning problems. The discipline is in the details: choose the window deliberately, respect seasonality, layer overrides on top rather than fiddling with the baseline, and always benchmark a more complex model against the moving-average baseline before claiming the complexity is worth the cost. The planner who knows when to reach for a moving average and when to reach past it is a more confident and more efficient forecaster than the one who treats simplicity as obsolete.

Pair this with exponential smoothing for the natural next step beyond moving averages, and the simple volume forecaster for a working example.