Exponential Smoothing

Forecasting by weighting recent history more than old history.

Also known as: exponential smoothing, Holt-Winters, ETS, simple exponential smoothing

Exponential smoothing forecasts by averaging history with weights that decay exponentially — yesterday matters more than last month, which matters more than last year. One parameter sets how fast the past is forgotten; extensions add trend and seasonality components (Holt-Winters), each with its own decay rate.

forecast = α·latest + (1-α)·previous forecast        (0 < α < 1)
α near 1:  chases every wiggle (nervous)    α near 0:  slow, stable (sluggish)

It is the pragmatic middle between a moving average and ARIMA: adaptive like the former (it tracks level shifts automatically), principled enough to carry trend and season, and simple enough to explain to stakeholders and maintain across thousands of series — which is why it still runs half the world’s inventory and demand forecasts.

The classic mistakes:

One α for every series. Fast-moving SKUs and slow staples need different memory. Fit per series (or per group), or the parameter is a compromise that suits nothing.
Smoothing through structural breaks. A price change or stockout enters the average and pollutes forecasts for months. Detect breaks and reset the level rather than letting the smoother digest them slowly.
No intervals. Like every point method, it needs an uncertainty band to be decision-grade — see forecast intervals.
Seasonality assumed, not checked. Holt-Winters with a seasonal component on non-seasonal data fits noise seasonally. Verify the season exists (seasonality) before paying for it.
Judging by fit, not holdout. Smoothing always looks good in-sample. Compare against naive on held-out trailing data (see forecast accuracy).

When not to use it: series driven by known external events (promotions, holidays as regressors beat pure smoothing), intermittent demand (needs specialised methods), or very short history (there is nothing to decay from — use naive or pooling).