Statistics Glossary → Damped trend
Damped trendconcept
A model component in exponential smoothing that makes the trend fade as the forecast horizon lengthens, so that the forecasts eventually level off.
In Holt's linear trend method, forecasts extend in a straight line as if the latest trend would last forever; over long horizons this often proves too optimistic or too pessimistic. With a damped trend (Gardner and McKenzie), the h-step-ahead forecast is ℓ_t + (φ + φ² + … + φ^h) · b_t, where 0 < φ < 1 is the damping parameter. At each step the contribution of the trend shrinks by a factor of φ, and the forecasts approach the level ℓ_t + φ b_t / (1 - φ).
φ = 1 corresponds to undamped Holt and φ = 0 to simple exponential smoothing without trend; in practice φ is usually estimated between 0.8 and 0.98. Empirical comparisons have often found damped trends to be more accurate for medium and long horizons, although they can hold back forecasts for series whose trend is known to continue. The 'damped stochastic cycle' of the unobserved components model is a different notion: there the damping affects the amplitude of a cyclical fluctuation, not the trend.
An everyday example
The number of users of a new app grows by 10,000 a month in its first months. A linear-trend forecast assumes this growth will continue unchanged for five years; a damped trend, allowing for the market becoming saturated, reduces the growth a little each month so that the forecast settles on a plateau over time.
In MerQur
Offered in Exponential Smoothing as the 'Damped trend (for long horizons)' option, together with 'Trend type' (Additive / Multiplicative).
Analyses where this concept appears
Exponential Smoothing (Holt-Winters)