What is time series forecasting?
Time series forecasting is the use of historical, time-ordered data to predict future values of a quantity. Demand for a product next month, call volume tomorrow, cash position at quarter end, and energy load this evening are all forecasting problems. The forecast horizon is how far ahead the prediction reaches, and the output is ideally a range with a stated confidence rather than a single number.
How time series forecasting works
Classical statistical methods such as exponential smoothing and ARIMA model trend and seasonality directly and remain hard to beat on short, clean series. Machine learning methods such as gradient-boosted trees use lag features, calendar features, and external drivers like promotions, weather, or price. Neural methods, including LSTM networks and transformer-based models, learn across many related series at once and suit large catalogs. Pre-trained forecasting models that work with little or no training data have also appeared.
Whatever the method, evaluation uses backtesting: the model is trained on data up to a past date, forecasts the following period, and the error is measured with metrics such as mean absolute error or mean absolute percent error. Every forecast should be compared with a naive baseline, such as the same period last year, because a surprising number of elaborate models fail to beat it.
Time series forecasting in practice
A food distributor forecasts weekly demand for each product at each depot. Many products sell intermittently, so the team uses methods designed for series with many zeros. Promotions and holidays enter as external features. Forecasts at product level are reconciled so they add up to the depot and regional totals that finance plans against. In Kirk Tech Solutions engagements, the first deliverable is usually a backtest of the proposed model against the planner’s current method on the company’s own history, before any system is built.
What to watch for
- Insist on a comparison with a naive baseline before accepting any accuracy claim.
- Match the forecast horizon to the decision lead time, since a perfect one-week forecast does not help a twelve-week supplier.
- Ask for prediction intervals, because inventory and staffing decisions depend on the range rather than the midpoint.
- Expect history to mislead after a structural change such as a new product line, a lost customer, or a pandemic.