Exponential smoothing & ETS

Exponential smoothing is the workhorse of supply chain forecasting because it is fast, needs little data and runs well across thousands of items. Every version rests on one idea:

new estimate = old estimate + α × (how wrong the last forecast was)

The smoothing constant α, between 0 and 1, sets how much the forecast reacts. A handy rule: the forecast effectively remembers about the last 1/α periods. With α = 0.1 it remembers about 10 periods and moves calmly; with α = 0.5 it remembers about 2 and reacts to every blip.

The family grows by adding pieces:

• Simple exponential smoothing (SES) tracks only the level, so its forecast is a flat line. Good for steady demand with no trend.
• Holt’s method adds a trend, so the forecast becomes a sloped line.
• Damped trend lets that slope fade over time. This matters, because an undamped trend keeps rising for ever: a level of 100 growing by 10 a month forecasts 220 a year out, while a damped version settles below 200. For longer horizons, damping is usually the safer default.
• Holt-Winters adds seasonality, either additive (the seasonal swing is a fixed amount) or multiplicative (the swing is a percentage, so it grows as the business grows, which suits most retail and FMCG data).

ETS (error, trend, seasonal) wraps all of these into one statistical framework. Software can then pick the best version automatically and give prediction intervals as well as a single number.

Simple exponential smoothing

The level $\ell$ is updated after each new observation $y_t$:

The forecast for every future period is the latest level. Unrolling the formula shows where the name comes from: past observations get weights $\alpha$, $\alpha(1-\alpha)$, $\alpha(1-\alpha)^2$ and so on, shrinking exponentially with age.

Adding trend: Holt’s method

Holt adds a slope $b$, smoothed with a second constant $\beta$:

The forecast $h$ periods ahead is $\ell_t + h\,b_t$, a straight line.

Damping the trend

A damped trend multiplies the slope by $\phi$ (just below 1) for each step ahead, so the forecast becomes $\ell_t + (\phi + \phi^2 + \dots + \phi^h)\,b_t$. With a level of 100, a trend of 10 a month and $\phi = 0.9$, the forecast 12 months ahead is about 165 instead of 220, and it can never exceed $100 + 10 \times 0.9/0.1 = 190$.

Adding seasonality: Holt-Winters

A third equation keeps a seasonal index for each period of the cycle (12 for monthly data), smoothed with $\gamma$. In the multiplicative form the forecast is $(\ell_t + h\,b_t) \times s$, where $s$ is the index of the season being forecast: 1.3 means 30% above a normal month. In the additive form the seasonal amount is added instead.

ETS: from recipes to statistical models

Exponential smoothing began as a set of practical recipes. The ETS framework recasts each one as a statistical model defined by three choices: the Error (additive or multiplicative), the Trend (none, additive or additive damped) and the Seasonality (none, additive or multiplicative). Each combination has a likelihood, which brings three benefits:

• Estimated parameters: $\alpha$, $\beta$, $\gamma$, $\phi$ and the starting values are chosen by maximum likelihood instead of by hand.
• Model selection: software fits the candidates and picks one with an information criterion (such as AICc) that rewards fit but penalizes complexity.
• Prediction intervals: the model gives a forecast distribution, not just a number, which is what safety stock and probabilistic planning need.

Reading the parameters

• $\alpha$ close to 1: the series behaves almost like a random walk, and the forecast is close to “same as last period”.
• $\alpha$ close to 0: a very stable level, with recent noise mostly ignored.
• $\beta$ close to 0: a steady trend; larger values let the slope change quickly.
• A tiny $\gamma$ means the seasonal shape barely changes from year to year.

Simple smoothing is equivalent to an ARIMA(0,1,1) model and Holt’s linear method to ARIMA(0,2,2), so the two families overlap more than their names suggest.

Why damped trend is a robust default

The damped trend method, introduced by Gardner and McKenzie in 1985, has been one of the most reliable performers on business data in forecasting competitions. Real trends rarely continue for ever, and a damped forecast hedges between “the trend continues” and “it stops”.

Running it across thousands of items

• Short histories: with less than two full seasons of data, seasonal models are unreliable. Borrow the seasonal shape from the product family.
• Promotions and stockouts: ETS has no inputs for drivers, so clean them out of the history to forecast the baseline, and add promotional uplift separately.
• Breaks: after a structural change, re-initialize or shorten the history; otherwise the model takes many periods to forget the old level.
• Weekly data: a year has about 52.18 weeks, which classic Holt-Winters can’t represent exactly. Models with Fourier terms handle long or non-integer seasons better.