Time-series forecasting is the task of predicting the future values of a quantity that is measured repeatedly over time, such as daily sales, hourly electricity demand, monthly unemployment, or a stock price. Because the data points are ordered in time and usually depend on their own recent past, forecasting needs methods that respect that structure rather than treating each observation as independent.
A typical time series can be thought of as a combination of a trend (the long-run direction), seasonality (patterns that repeat over fixed periods like weeks or years), and irregular noise; many methods work by modeling these pieces. The toolkit ranges from simple classical methods such as exponential smoothing and ARIMA, through decomposition techniques like STL, to machine learning approaches including gradient-boosted trees, deep networks such as DeepAR and N-BEATS, and the recent time-series foundation models. A central lesson from the M-competitions, the field’s main benchmark, is that no single method wins everywhere: the best choice depends on the data, and combinations of methods often beat any one alone. Good practice also means forecasting uncertainty, a range of likely outcomes, not just a single number, since plans must account for risk.
Forecasting is among the most economically important applications of statistics and machine learning.
Why business readers should care: nearly every operational plan, from inventory to staffing to budgets, rests on a forecast, so the quality of those forecasts directly affects cost and service.