Skip to content
ML Academy

Introduced on day 23

Exponential smoothing

Forecasting by weighted averages of past observations whose weights decay geometrically, with separate smoothing for level, trend and season.

In detail

Simple exponential smoothing updates a level with parameter alpha; Holt adds a trend with beta and Holt-Winters a season with gamma. The ETS framework writes these as innovations state-space models with additive or multiplicative error, trend and season, which provides a likelihood, AICc selection and prediction intervals. High smoothing parameters make the model chase recent values.