MerQuR+Statistics Glossary

Statistics Glossary → Exponential smoothing parameters (α, β, γ, φ)

Exponential smoothing parameters (α, β, γ, φ)concept

The parameters in exponential smoothing methods that set how quickly the level, trend and seasonal component of a series are updated with new observations, and how strongly the trend is damped.

Simple exponential smoothing (SES) tracks only the level: new level = α·(new observation) + (1 − α)·(old level), so the weights of past observations decay exponentially going back in time. With α close to 1 the forecast is very sensitive to the latest observation; close to 0, it stays near the average of the long past. Holt's method adds β for the trend, and Holt-Winters adds γ for the seasonal component. With a damped trend, φ (0 < φ < 1) makes the trend flatten out gradually in long-range forecasts. In the state-space (ETS) formulation the trend parameter is often written β*, related to the classical β by β = α·β*.

The parameters are usually estimated from the data by minimising forecast errors or maximising the likelihood, and the choice between models (e.g. SES, Holt, damped Holt, Holt-Winters) is made with criteria such as AICc. An estimated parameter very close to zero means the corresponding component is almost constant; γ ≈ 0, for example, shows that the seasonal pattern does not change over the years. A damped trend generally reduces the overly optimistic or pessimistic results of a linear trend at long horizons. Fixing parameters by hand should be done only with a theoretical reason or for comparison.

An everyday example

A bakery forecasts daily bread sales. With α = 0.8, yesterday's unusually high sales almost entirely determine tomorrow's forecast; with α = 0.2, the forecast stays close to the general level of the past few weeks and is little affected by a single day's fluctuation. If sales change slowly, a small α gives more stable and usually more accurate forecasts.

In MerQur

In exponential smoothing, the 'Method' option chooses between the Automatic, Holt-Winters, Holt, SES and Simple seasonal models, with automatic selection by the smallest AICc. 'Fixed parameters (α, β*, γ, φ)' fixes the chosen parameter while the others and the initial states are estimated.

Analyses where this concept appears

Exponential Smoothing (Holt-Winters)

Related concepts

Damped trendSeasonality and seasonal period