Exponential Smoothing: Weighting Recent Data More Heavily

The mental model

Exponential smoothing creates forecasts by calculating a weighted average of past observations. Unlike a simple moving average where all points have equal weight, it gives exponentially more weight to recent data. The influence of an observation decreases as it gets older, which is often a good assumption for real-world systems where recent behavior is the best predictor of the near future.

How it works

Simple Exponential Smoothing (SES) uses a single parameter, alpha (α), called the smoothing factor, with a value between 0 and 1. The forecast for the next time step is a combination of the most recent observation and the most recent forecast. The formula is: Next Forecast = α (Current Observation) + (1 - α) (Previous Forecast). A high α makes the model react quickly to recent changes, while a low α results in a smoother forecast that is less sensitive to recent fluctuations. More advanced methods like Holt's and Holt-Winters build on this by adding parameters for trend and seasonality.

When to use it

Use exponential smoothing for generating reliable forecasts quickly and for a wide range of time series, especially when you need to automate forecasting for many items at once. It's common in industry for tasks like inventory management (predicting demand for thousands of SKUs), monitoring server metrics, or short-term financial forecasting. Its simplicity and speed are major advantages.

When not to use it

Avoid the simple version of exponential smoothing for data with a clear trend or seasonality. Because SES assumes the data fluctuates around a stable mean, it will consistently lag behind a growing series or fail to predict cyclical peaks and troughs. For such cases, you must use more complex variations (like Holt-Winters) or different models entirely. It also cannot account for external causal factors, like a marketing promotion's effect on sales.

One canonical example

Imagine forecasting daily app users. Your forecast for today was 5,000 users, but the actual number was 5,500. Using a smoothing factor α of 0.2, your forecast for tomorrow would be: (0.2 5,500) + (1 - 0.2) 5,000 = 1,100 + 4,000 = 5,100. The new forecast is adjusted slightly upward from the old forecast, influenced by the recent higher-than-expected value.