Does time series forecasting require deep learning?

No. Simple benchmarks, regression, exponential smoothing and other statistical methods can be useful.

Can forecasts use outside information?

Yes, provided the information is available at the forecast origin or its future uncertainty is explicitly handled.

Is the longest horizon always the hardest?

Often uncertainty grows with distance, but seasonal structure and the particular target can create exceptions; evaluate the horizons actually needed.

AI Glossary / Time series and signal processing / Beginner

Time Series Forecasting

Definition

Time series forecasting estimates future values of a quantity from information available at a specified time. It uses the ordering of observations, and sometimes additional predictors, to anticipate what may happen over a chosen horizon. The output can be a point estimate, an interval, or a probability distribution. Forecasting supports decisions under uncertainty; it does not make future observations certain or require a particular AI architecture.

What it means

A forecast answers a time-specific question, such as how many requests may arrive next week based on what was known this Friday. That differs from describing last week's pattern or filling a missing historical value. The forecast origin is the information cutoff, and the horizon identifies how far ahead the prediction reaches.

Historical repetition can be informative, but its usefulness depends on whether relevant relationships persist. A holiday calendar may be known ahead of time, while next month's realized weather is not. Using the latter during historical testing without accounting for its availability makes a forecasting system appear better informed than it would have been in practice.

How it works

Define the target, observation frequency, forecast origin and operational horizon. Examine missing periods, changes in measurement, trend and seasonality. Establish simple benchmarks before testing more elaborate statistical or machine-learning methods. Train on data available before each evaluation origin and generate predictions for later periods.

Compare forecasts with subsequently observed outcomes using measures suited to the decision. Evaluate uncertainty as well as central estimates, and inspect performance by horizon and relevant operating conditions. Updating a model with newly available observations is legitimate, but every historical evaluation must reproduce when those observations actually became available. Keep issued forecast versions so later assessment does not accidentally compare revised predictions with earlier decisions.

Three practical examples

Staffing: forecast hourly service requests several days ahead so managers can compare possible staffing levels with expected demand and uncertainty.
Inventory: forecast weekly product demand over a replenishment lead time, while distinguishing recorded sales from demand lost during stockouts.
Household planning: estimate electricity consumption for the coming week using past meter readings and information genuinely available when the estimate is made.

Think of it this way

Think of preparing for a journey using the route already traveled, recurring traffic patterns and the information currently available. Those clues help anticipate what lies ahead, but unexpected roadworks can change the outcome. The analogy emphasizes information and uncertainty; a forecasting method must express its numerical assumptions more precisely than a traveler's informal expectations.

Worked example

Suppose an invented quarterly demand series ends with the most recent year's values 120, 150, 200 and 180 units. At the end of that year, a seasonal naive benchmark forecasts the next four quarters as 120, 150, 200 and 180. A nonseasonal naive benchmark repeats the latest observation, giving 180 for every future quarter. Neither method uses observations from the coming year.

Later, actual demand is 125, 148, 205 and 185. The seasonal forecast has absolute errors 5, 2, 5 and 5, giving mean absolute error 17/4 = 4.25 units. The nonseasonal forecast has absolute errors 55, 32, 25 and 5, giving 117/4 = 29.25 units. The seasonal benchmark performs better on these four cases, but one year is weak evidence for a universal choice.

All four predictions were issued from the same origin. Replacing later predictions with forecasts updated after each actual quarter arrives would create a different evaluation. Both designs can be useful, but they answer different planning questions. No prediction interval is supplied here because these eight values alone do not justify a particular uncertainty model.

Illustrative case study

A fictional community venue forecasts attendance to plan volunteer shifts. Its first model uses final event-registration totals, although those totals are not known when shifts must be scheduled. The team rebuilds its test using registration snapshots from the actual planning deadline. It compares the revised model with a seasonal benchmark and examines unusually busy events separately. The resulting forecast is less flattering but more relevant to the scheduling decision because its information matches the real workflow.

Common misconceptions and mistakes

Forecasting is not simply fitting a curve through historical points. A model can reproduce the past and fail on future periods. Randomly splitting neighboring observations can leak information or evaluate the wrong deployment situation. A point forecast is not a guaranteed outcome, and a narrow interval is not automatically a trustworthy one.

More history is not always better when definitions or behavior have changed. Forecast error also does not establish its cause: unusual demand, incomplete recording and model limitations can produce similar discrepancies. Forecasts conditional on a planned promotion should not be interpreted as proof that the promotion causes the predicted increase.

Business use

Connect the forecast horizon to a real decision deadline. A next-day estimate may be accurate yet arrive too late for a supplier requiring two weeks' notice. Record the target units, data cutoff, assumptions about future predictors and acceptable consequences of overprediction or underprediction.

Compare the whole planning process with a practical baseline. Review forecast revisions, missed events and uncertainty coverage after outcomes mature. Where different decisions require different risk tolerances, a probability distribution or selected quantiles may be more useful than one shared central estimate. Forecast quality and decision quality remain related but distinct assessments.

Check your understanding

Why is it misleading to test a Friday forecast using next week's realized weather?

Answer: That weather was unavailable on Friday. The test would give the model information beyond its stated cutoff unless it explicitly evaluates a conditional scenario instead of the operational forecasting process.

Related terms

Time series
Forecast horizon
Prediction interval
Mean Absolute Error
forecast origin
seasonal naive forecast
temporal validation
trend
seasonality
forecast revision

Learn next

Start by plotting one regularly observed quantity and writing down the exact decision its forecast should support. Produce a last-value and a seasonal benchmark. Then replay several historical forecast origins, checking information availability and errors by horizon before introducing a more complicated model.

Sources

Hyndman and Athanasopoulos, Forecasting: Principles and Practice, forecaster's toolbox and benchmark methods: https://otexts.com/fpp3/toolbox.html Hyndman and Athanasopoulos, dynamic regression and future predictor availability: https://otexts.com/fpp3/dynamic-regression.html Scope: general forecasting task; the numerical comparison and venue case are original illustrations, not reported research outcomes.

Previous: Time series breaks Next: Time series model