Can forecasting be done without much historical data?

Structured judgment and information from related situations can help, but the evidence limitations should be explicit.

Is every prediction a forecast?

Usage varies, but forecasting usually emphasizes future outcomes; predicting a missing past label is a different task.

Can forecasts influence outcomes?

Yes. Decisions based on forecasts can change demand, capacity, or behavior, which complicates evaluation and may require explicit modeling of those actions.

AI Glossary / Time series and signal processing / Intermediate

Forecasting

Definition

Forecasting is the process of estimating future outcomes using information available at the time the forecast is made. It can combine historical observations, statistical models, machine learning, known future events, and structured expert judgment. A forecast may be a single predicted value, a range, a probability distribution, or probabilities for future events.

What it means

A forecast describes what may happen under stated information and assumptions. It is different from a goal, which describes what someone wants to happen, and a plan, which specifies actions intended to influence outcomes.
Forecasting does not require certainty to be useful. A realistic estimate of several possible outcomes can improve preparation more than an unjustifiably precise number. Its value depends on whether the prediction arrives in time and addresses the quantity relevant to the decision.

How it works

Define the target, unit, time frequency, forecast horizon, and information cutoff. Assemble relevant data, examine quality and patterns, and establish simple benchmarks. Develop candidate methods using only information that would have been available at each forecast origin.
Evaluate predictions on later observations, ideally across multiple historical origins that resemble deployment. Assess uncertainty as well as point accuracy, select a method appropriate to the decision, and monitor subsequent performance. Document any human adjustments and their rationale. When external predictors are needed, account for whether their future values are known, separately forecast, or assumed in scenarios.

Three practical examples

A retailer estimates demand during a replenishment lead time, then uses the forecast distribution and stockout costs to inform ordering.
A service team predicts incoming requests before staffing schedules are fixed. It evaluates performance at that same scheduling horizon.
A household forecasts upcoming expenses from recurring commitments and variable spending patterns, revising the estimate when new information becomes available.

Think of it this way

Think of preparing for an outdoor event using a weather outlook. The forecast describes possible conditions; the goal may be a successful event; the plan may include a shelter or an indoor alternative. These are connected activities, but changing the desired outcome does not by itself change the evidence about future conditions.

Worked example

Consider an original evaluation of forecasts for four separate future weeks. A simple benchmark issued before each week predicts 100, 120, 110, and 130 service requests. A candidate method, also issued before each target week, predicts 105, 122, 116, and 134. The subsequently observed counts are 108, 125, 119, and 138.
Using actual minus forecast, benchmark errors are 8, 5, 9, and 8. Its mean absolute error is (8 + 5 + 9 + 8)/4 = 7.5 requests. Candidate errors are 3, 3, 3, and 4, giving mean absolute error 13/4 = 3.25 requests. The relative reduction in this metric is (7.5 − 3.25)/7.5, approximately 56.67%.
Every error is positive, so both methods underforecast every week in this tiny sample. The candidate performs better on these four cases, but four observations do not establish a durable advantage, uncertainty calibration, or statistical significance. No prediction interval can be responsibly inferred solely from the two error averages. The calculations illustrate a transparent benchmark comparison, not a proven forecasting product.

Illustrative case study

Illustrative fictional case: Larch Services initially asks its analyst to forecast next month's demand at the sales target. The analyst separates the target from a data-based forecast and presents the gap explicitly.
Managers then develop a marketing plan and request a conditional scenario that includes the proposed activity. They retain the original forecast, record the scenario assumptions, and compare later outcomes with both. This fictional case demonstrates how forecasting can inform action while preserving a clear record of what was expected before the action was chosen.

Common misconceptions and mistakes

A good historical fit is not the same as forecasting accuracy. A model can reproduce past observations while failing on future ones, especially if tuning or preprocessing used later information. More complexity, more data, or an AI label does not guarantee better forecasts.
Do not present a conditional scenario as an unconditional prediction. A forecast assuming a promotion occurs is different from one that averages over whether it occurs. Nor does predictive success establish a causal relationship: a useful indicator may not be an effective intervention. Forecasts should also distinguish observed sales from latent demand when stockouts constrain what can be sold.

Business use

Forecasting helps organizations allocate resources, coordinate operations, and prepare for uncertainty. A useful system connects forecast outputs to actual decisions and compares the costs of different errors. Underprediction and overprediction can have different consequences even when their absolute sizes match.
Maintain versioned issue dates, assumptions, and evaluation results. Compare against credible simple baselines and monitor changes in the operating environment. Where human judgment is added, measure its incremental contribution rather than assuming that every adjustment improves the model.

Check your understanding

Why must the example's forecasts be issued before the target weeks?

Answer: Otherwise the comparison could include hindsight rather than genuine prediction.

Does lower mean absolute error prove that uncertainty intervals are reliable?

Answer: No. Interval calibration requires separate evaluation.

What does a positive actual-minus-forecast error mean?

Answer: The outcome exceeded the forecast.

Is a sales target of 200 evidence that next month's sales will be 200?

Answer: No. It is a desired outcome unless supported by a separate forecasting argument.

Related terms

Forecast horizon
Prediction interval
Benchmarks
Calibration
forecast origin
point forecast
forecast distribution
scenario analysis
benchmark
rolling-origin evaluation
forecast error

Learn next

define one practical forecasting decision, including when action must be taken. Build a simple benchmark and evaluate it on genuinely later outcomes before adding complexity. Then connect forecast uncertainty to the decision's costs and available responses.

Sources

Sources: Hyndman and Athanasopoulos, Forecasting: Principles and Practice, third edition, Getting started: https://otexts.com/fpp3/intro.html . Glossary of Forecasting Terms, Forecasting: https://robjhyndman.com/mwh3/FG4.pdf . Primary textbook and supplied forecasting glossary consulted. The benchmark calculations and fictional case are original; no real organizational performance is claimed.

Previous: Forecast skill Next: Foreign key