Exponential Smoothing
Exponential smoothing forecasts the next value as a weighted blend of the latest observation and the previous forecast. Recent data counts more, and older data fades smoothly.
The smoothing factor
A single parameter called alpha between zero and one controls the blend.
- A high alpha trusts the newest observation and reacts quickly.
- A low alpha leans on history and produces a calmer, slower forecast.
Each forecast equals alpha times the current value plus one minus alpha times the prior forecast. Expanding this shows the weights decay exponentially into the past.
Why use it
- It needs only the last forecast and the last value, so it is cheap and easy to update online.
- It adapts to gradual shifts without storing a long history.
Limits
Plain exponential smoothing assumes no trend and no season. It will lag a rising series and ignore cycles. Extensions add a trend term and a seasonal term to fix this, which leads toward the Holt and Holt Winters methods.
Key idea
Exponential smoothing blends the latest value with the prior forecast, giving recent data exponentially more weight through a single factor.