Forecasting method: exponential smoothing and Holt-Winters

Forecasting · ~7 minute read

The forecasting workhorse

Exponential smoothing is the most consistently useful forecasting family in contact centre planning. It is simple enough to explain to finance in three minutes, flexible enough to model level, trend, and seasonality, and competitive enough on accuracy that the latest machine-learning methods often struggle to beat it by meaningful margins. Almost every WFM platform has it under the hood. Many planners use it daily without naming it explicitly. This article walks through the three variants — single, double, and triple (Holt-Winters) — explains what the smoothing parameters do, and sets out where the method earns its place and where it does not.

The basic idea

Exponential smoothing forecasts the future as a weighted average of past observations, with the weights declining exponentially as you go back in time. The most recent observation has the most influence; the observation a year ago has very little. The rate of decline is controlled by a smoothing parameter, usually called alpha, between zero and one. A high alpha (say 0.7) makes the forecast adapt quickly to recent change; a low alpha (say 0.1) makes it smooth out noise and respond slowly. The planner’s job is to pick a value of alpha that fits the data, or to let the software fit it automatically by minimising forecast error on historical data.

Single exponential smoothing

Single exponential smoothing forecasts the next period as a smoothed version of the recent observations. It assumes there is no trend (the underlying level is roughly constant) and no seasonality. This is rare in contact centre data — most queues have some pattern — but single exponential smoothing is useful for back-office work, very stable queues, and as a baseline for testing more complex methods. The forecast is simple: next period equals alpha times the latest observation plus (1 minus alpha) times the previous forecast. The model has one parameter, fits in seconds, and adapts to slowly-changing data gracefully.

Double exponential smoothing (Holt)

Double exponential smoothing — often called Holt’s method — adds a trend component. There are now two smoothing parameters: alpha for the level (as before) and beta for the trend. The forecast at any future point is the current level plus the trend multiplied by the number of periods ahead. This makes the method useful for queues with a steady underlying growth or decline that the simple moving average would lag behind. The trade-off is that the model is more sensitive to noise that looks like trend — a couple of unusually busy days can convince a high-beta model that the operation is growing fast, and the forecast then over-shoots. Lower beta values (0.05 to 0.15) tend to produce more robust forecasts in practice.

Triple exponential smoothing (Holt-Winters)

Triple exponential smoothing — the Holt-Winters method — adds a seasonal component. There are now three smoothing parameters: alpha for level, beta for trend, and gamma for seasonality. The seasonal period is set in advance (seven for daily data with weekly seasonality, twelve for monthly data with annual seasonality, and so on). The forecast for a future period is the current level, plus the trend extrapolated forward, plus the appropriate seasonal factor for that period.

Holt-Winters comes in two flavours: additive, where the seasonal effect is added to the level (a Monday gets +30 contacts more than the average day), and multiplicative, where the seasonal effect is a multiplier (a Monday gets 1.15 times the average day). Multiplicative is the right choice when the seasonal amplitude grows with the level — busy quarters have bigger Monday peaks than quiet quarters. Additive is the right choice when the amplitude is roughly constant. The simple volume forecaster on this site uses additive Holt-Winters as its default.

How to choose the smoothing parameters

Three approaches consistently work. The first is auto-fit: search a grid of alpha, beta, and gamma values, fit the model to historical data, and pick the combination that minimises in-sample forecast error. Almost every statistical package does this automatically. The second is operationally-informed: pick conservative values (alpha around 0.2–0.3, beta around 0.05–0.1, gamma around 0.2–0.4) and adjust only if the model is clearly drifting. The third is to combine them: auto-fit gives you a starting point, then the planner adjusts on operational judgement (a high alpha after a known step-change in volume, for example).

Where Holt-Winters earns its keep

Holt-Winters is the right default for most contact centre forecasting problems. It captures the three components every planner cares about — baseline level, trend, and seasonality — with three parameters that are intuitive enough to explain and adjust. It runs in milliseconds, so it can be re-fitted weekly without infrastructure cost. It is robust: missing the parameters by a reasonable margin still produces a reasonable forecast. And it is well-understood enough that any planner can explain it in detail to finance, audit, or operations without resorting to hand-waving.

Where it falls short

Three situations push the planner past Holt-Winters. The first is multiple overlapping seasonalities — a queue with both a weekly pattern and a monthly billing cycle, for example. Classical Holt-Winters handles one seasonal period at a time; multiple periods need either more sophisticated state-space models or explicit feature engineering with regression. The second is strong dependence on external drivers — weather, marketing campaigns, system events. Holt-Winters has no way to use these; the planner has to layer manual overrides on top, or move to a regression-based method that includes them. The third is structural change — the method assumes the underlying patterns are stable, and adapts to change only as fast as alpha and beta allow. A pandemic, a merger, or a sudden product launch can break the model for months.

Common mistakes

Three patterns recur. Setting the seasonal period wrong (using 12 when the data is actually weekly, for example), which produces nonsensical forecasts. Letting auto-fit choose extreme parameters that fit historical noise rather than signal — an alpha of 0.95 says “trust the latest observation almost entirely,” which is rarely the right answer. And re-fitting too often, which makes the forecast jumpy from week to week without improving accuracy — monthly or quarterly re-fitting is usually plenty.

Conclusion

Holt-Winters is the single most useful forecasting method in contact centre planning. It is fast, transparent, well-understood, and competitive with much more complex approaches on most queues. A planning function that uses it well as a baseline and reaches past it only when the data genuinely justifies the move has a defensible methodology, a fast workflow, and a forecast it can explain to anyone in the operation. The harder skill is knowing when not to use it — when external drivers, multiple seasonalities, or structural change have made the assumptions invalid — and that judgement is what separates competent forecasting from excellent forecasting.

Pair this with moving averages for the foundation, ARIMA for the next step, and the simple volume forecaster for a working Holt-Winters implementation you can experiment with.