Exponential smoothing is a family of forecasting methods built on weighted averages of the past.
It exists because recent values usually say more about the next value than distant ones.
It is fast, needs little data and adapts as new values arrive.
Simple exponential smoothing keeps a single level, updated with each new value.
The new level is the smoothing factor times the latest value, plus 1 minus the factor times the old level.
For example, with a factor of 0.3, a level of 100 and a new value of 120, the new level is 106.
The forecast for every future period is that latest level, which suits series with no trend or season.
Holt's method adds a trend that is smoothed in the same way, so forecasts slope upwards or downwards.
Damped trends flatten the slope over the horizon, which often forecasts better than a straight line.
Holt-Winters adds a seasonal part, updated each period, for series with weekly or yearly patterns.
The smoothing factors are chosen by making past forecast errors as small as possible.
Versions known as ETS models add prediction intervals to the forecasts.
Exponential smoothing struggles with sudden breaks, such as a pandemic, which no weighting of the past predicts.
In practice it is a common default for forecasting demand across thousands of products.
The usual mistake is skipping simple baselines like this one in favour of complex models that forecast no better.