Glossary

Exponential smoothing

A forecasting technique that weights recent observations more heavily than older ones, with the weight decaying exponentially with age.

Also called: ETS model

Exponential smoothing forecasts a time series by taking a weighted average of past observations in which the weight given to each point decays exponentially the further back it lies, so the most recent observation matters most and influence fades gradually rather than dropping off sharply at a fixed cutoff. Simple exponential smoothing produces a flat forecast; extended versions add trend and seasonal components to handle series that are rising, falling, or cyclical.

The simplest form updates a level estimate as forecast = alpha * latest_value + (1 - alpha) * previous_forecast, where alpha, between 0 and 1, controls how quickly the forecast reacts to new data. This differs from a simple moving average, which weights a fixed window of recent points equally and drops older points abruptly once they leave the window, rather than fading them out gradually. Holt-Winters exponential smoothing extends this with separate trend and seasonality terms, making it a common alternative to ARIMA for series with clear seasonal patterns.

Exponential smoothing matters because it is simple to implement, computationally cheap, and performs competitively on many business time series without the parameter search ARIMA requires, which is why it remains a common default in automated time series forecasting tools. A common pitfall is picking a smoothing parameter that is too reactive, producing a forecast that chases noise, or too slow, producing one that lags real shifts in the underlying trend.

Last reviewed September 22, 2026

In the index now

Related terms

Moving average A calculated series that averages a fixed window of recent data points, smoothing out short-term noise to reveal a trend. ARIMA (autoregressive integrated moving average) A classical time-series forecasting model that combines a series' own past values, differencing, and past forecast errors. Time series forecasting Predicting future values of a metric using patterns found in its own historical sequence over time. Seasonality A regular, predictable pattern in data that repeats at a fixed calendar interval, such as daily, weekly, or annually. Weighted average An average in which some values contribute more to the result than others, based on an assigned weight.

Related tools

Darts Open-source Python library offering one unified API across dozens of classical, ML and deep-learning forecasting models.

Related guides

Time series analysis, explained How analysts model data ordered in time — the patterns forecasting methods look for, the pitfalls, and where the data actually lives.