Time series forecasting

Time series forecasting predicts future values of a metric by modeling patterns in its own history, ordered over time—trend (long-term direction), seasonality (repeating cycles like day-of-week or month-of-year effects), and shorter-term autocorrelation between consecutive periods. Unlike general regression analysis, the key predictor is the variable's own past, not a separate set of external features, though modern approaches often blend both.

Common methods range from classical statistical models (moving averages, exponential smoothing, ARIMA) to machine learning approaches that can incorporate external drivers like price, weather, or promotions alongside historical patterns. Forecasts are evaluated by holding out recent periods and measuring how far predictions land from actual outcomes, and should always be reported with an uncertainty range rather than a single number, since uncertainty typically grows the further out the forecast extends.

Time series forecasting underpins sales forecasting, demand forecasting, staffing plans, and capacity planning—anywhere a business needs to plan resources ahead of demand. It also plays a role in anomaly detection, since a forecast gives a baseline for flagging unexpected deviations. The main pitfalls are assuming past seasonality will hold after a structural change (a new product line, a sudden shift in customer behavior) and ignoring forecast uncertainty by presenting a single point forecast as if it were certain.