Classical Forecasting
Moving Averages & Exponential Smoothing
9 min+80 XPStep 1/10
Recent history should vote loudest
A moving average forecasts with the mean of the last k observations — every point votes equally, and the window slides forward. It smooths noise beautifully but carries a fatal habit: it always lags a trend, because stale points drag the forecast down.
Exponential smoothing fixes the weighting instead of the window. Each older observation's weight decays by a constant factor — a geometric fade controlled by α (between 0 and 1):
αnear 1: the forecast tracks recent data aggressively — responsive but jitteryαnear 0: long memory and smooth forecasts — stable but slow to react
In practice α is fit to minimize historical forecast error: the model chooses its own memory length.
analogy
A moving average is a committee where the last k days vote equally. Exponential smoothing is a committee where yesterday shouts, last month whispers, and last year is barely in the room.