In Holt-Winters, when do you choose multiplicative rather than additive seasonality?

## The three states Holt-Winters extends exponential smoothing with two extra pieces of state beyond the level: a **trend** (the per-period slope) and a **seasonal component** for each of the m positions in the cycle — m = 12 for monthly data with a yearly pattern, 7 for daily data with a weekly pattern, and so on. Each state gets its own smoothing parameter: alpha for the level, beta for the trend, gamma for the seasonal component. All three are updated every period from the newest observation and then combined into a forecast. The two seasonal forms differ in **how the seasonal state enters the forecast**. **Additive:** ``` forecast_{t+h} = level_t + h * trend_t + seasonal_{t+h-m} ``` The seasonal component is measured in the units of the series and is simply added on. December might be +4,000 units above the year's level, in every year, regardless of whether the level is 10,000 or 100,000. **Multiplicative:** ``` forecast_{t+h} = (level_t + h * trend_t) * seasonal_{t+h-m} ``` The seasonal component is a dimensionless factor. December might be 1.9 times the year's typical month — so when the level is 10,000 the December bump is 9,000 units, and when the level is 100,000 the same factor produces a 90,000-unit bump. By construction, additive seasonal components sum to roughly zero across one full cycle, while multiplicative factors average roughly one (they sum to about m). ## The decision rule Ask one question: **does the seasonal swing scale with the level?** - If the peaks and troughs keep the same absolute distance from the trend line as the series rises, the effect is additive. Heating-degree-driven quantities and many count series with a stable base behave this way. - If the peaks and troughs fan out as the series rises — the picture widens like a megaphone — the effect is multiplicative. Monthly champagne sales are the standard multiplicative example. The December spike is not a fixed number of extra bottles; it is roughly a fixed proportion of how much the business sells in an ordinary month. Over a few years of growth, the absolute size of the December spike grows with the business. Fit an additive model to that and the model must compromise: it settles on an average absolute December bump, then under-forecasts the peak in the strong later years and over-forecasts it in the weaker early ones. The residuals give this away — they show seasonal structure whose amplitude tracks the level rather than looking like noise. ## Practical constraints **Positivity.** Multiplicative seasonality multiplies the level by a factor, so it is only defined for strictly positive series. A series containing zeros or negative values (net changes, profit, temperature in Celsius) forces the additive form. A series that merely approaches zero is also dangerous: seasonal factors become unstable when the level they are dividing by is near zero during estimation. **Enough cycles.** Both forms need at least one full cycle of data to initialise the m seasonal states, and in practice at least two so the estimates are not pinned to a single year's idiosyncrasies. With eighteen months of monthly data you can technically fit a seasonal model, but you should say out loud that the seasonal estimates are weak. **A separate choice from the trend.** The trend can be absent, linear or damped independently of the seasonal form, and the error can be additive or multiplicative independently again. The ETS state-space naming captures all of this as a triple — error, trend, season, each none / additive / multiplicative with the trend optionally damped. Additive Holt-Winters is ETS(A,A,A) and the multiplicative-seasonality version with additive error is ETS(A,A,M). Naming the model this way is what lets you estimate it by likelihood, compare candidate forms by information criterion, and produce prediction intervals rather than just point forecasts. **Error form matters for intervals.** A series whose seasonal amplitude grows with the level usually has noise that grows with the level too. Choosing multiplicative seasonality but leaving the error additive fixes the point forecasts and still leaves prediction intervals of constant width at all levels — too wide at the bottom of the range and too narrow at the top. A fully multiplicative-error form addresses that. ## How to decide when the plot is ambiguous Plot first: the seasonal-amplitude-versus-level relationship is usually visible. If it is genuinely borderline — a short series, or a level that has not moved much — fit both forms within the ETS family on the same data and compare them by a corrected information criterion, then check the residuals of the winner for leftover seasonal structure. If the level barely varies over the sample, the two forms will fit almost identically, and the honest answer is that the data cannot distinguish them; pick the one that will still be correct if the level moves in the direction you expect. ## Common mistakes Choosing multiplicative because the series *trends* upward (trend and seasonal amplitude are separate questions), applying multiplicative seasonality to a series containing zeros, or treating the choice as cosmetic. It is not cosmetic: at a level far from the sample average, the two forms give materially different forecasts.