Can an ex ante forecast use future holidays?

Yes, when the calendar information is already known.

Can it use realized future weather?

Not as information supposedly available at the earlier origin.

Is an ex post forecast useless?

No; it can help diagnose sources of forecast error.

AI Glossary / Time series and signal processing / Intermediate

Ex ante forecast

Definition

An ex ante forecast is a forecast produced using only information available at the time it is made. It respects the forecast origin: the point at which the forecaster must act without knowing future outcomes. The term distinguishes genuine advance prediction from retrospective calculations that use information learned later.

What it means

A forecast can look accurate in a historical spreadsheet because the spreadsheet contains values that were unavailable when the decision had to be made. Future weather, revised economic data, and final campaign spending are common examples.
An ex ante forecast asks what could actually have been predicted at the time. This is essential when evaluating whether a forecasting system would have helped a real operational decision rather than merely explained the past after the fact.

How it works

Choose a precise forecast origin and horizon. Record when the forecast was issued, which future quantity it predicts, and when that quantity will be evaluated. An issue time can matter as much as a calendar date when information arrives during the day.
Define the available information set. A variable's reference period and publication date are different: a statistic describing last month may not have been released by today's forecast origin. Historical data revisions can also change values that were originally available.
Fit the model using only permitted information. This includes preprocessing, feature selection, parameter tuning, and any human adjustments. A model trained on later outcomes is not an ex ante reconstruction merely because its final input row has an earlier timestamp.
Supply future predictor values appropriately. Calendar features may already be known. Unknown predictors need forecasts, planned values, or explicitly labeled scenarios. The actual future realization cannot be substituted while retaining the claim of a genuine advance forecast.
Distinguish an ex post calculation. In regression forecasting, it can be useful to recompute a forecast with subsequently observed predictor values while holding the earlier-fitted model fixed. This helps diagnose predictor uncertainty, but it is not what the forecaster knew originally.
Preserve forecast vintages. Store the original forecast, model version, input snapshot, predictor forecasts, and relevant assumptions. Later updates should create new versions rather than overwrite the record used for evaluation.
Evaluate with rolling origins or another time-respecting design that mirrors deployment. Reconstruct data availability as closely as possible. A historical simulation may be called pseudo-out-of-sample, but its credibility depends on whether it uses realistic vintages and avoids later information.
Represent uncertainty consistently. An interval conditional on an assumed future predictor path omits uncertainty about that path. An operational ex ante forecast may need to propagate predictor uncertainty as well as future outcome noise and parameter uncertainty.

Three practical examples

A retailer forecasts tomorrow's demand using today's weather forecast rather than tomorrow's observed weather.
A capacity planner uses the campaign schedule approved at the forecast origin, recording later schedule changes separately.
A research team reconstructs historical forecasts using data releases that were actually available at each issue date.

Think of it this way

Think of predicting a match before kickoff with the information then available. Recalculating after learning the final lineup or weather can explain why the original prediction missed, but it must not replace the original prediction in the record.

Worked example

Consider an original fictional demand model fitted before a forecast is issued: predicted orders = 200 + 15 times temperature. At the issue time, tomorrow's temperature forecast is 24 degrees, so the ex ante order forecast is 200 + 15 x 24 = 560.
Tomorrow's actual temperature later turns out to be 28 degrees, and actual orders are 635. The original forecast error, actual minus forecast, is 635 - 560 = 75.
Holding the earlier-fitted coefficients fixed but inserting actual temperature gives an ex post prediction of 200 + 15 x 28 = 620. Its error is 635 - 620 = 15. The numerical difference of 60 equals 15 x (28 - 24), showing how the predictor realization changes this linear model's prediction.
This is an algebraic decomposition for the stipulated example. It does not prove that weather caused exactly 60 additional orders or that all remaining error has one source. Reporting 15 as the original operational forecast error would incorrectly use information unavailable at issue time.

Illustrative case study

Fictional case: Cedar Coast Events evaluates an attendance forecast using a historical weather file. The scores look unusually strong. A reviewer discovers that the file contains actual event-day weather rather than the forecasts available when staffing was scheduled.
The team rebuilds the evaluation using archived weather forecasts where available and labels periods without suitable archives as limited reconstructions. It keeps the ex post calculation as a diagnostic comparison rather than presenting it as deployment performance.

Common misconceptions and mistakes

Ex ante does not mean unbiased, accurate, or certain. A forecast can use assumptions and still be ex ante if those assumptions were available and documented at issue time. A train-test date split alone does not prevent leakage through revisions or preprocessing. A known calendar predictor is different from an unknown future measurement. Updating a forecast later creates a new forecast origin, not a correction to what was originally knowable.

Business use

Ex ante evaluation helps businesses judge forecasting systems fairly against the conditions under which decisions are made. Maintain issue timestamps and data vintages, and compare models using the same information set and horizon. This prevents an apparently superior model from winning simply because it received later or better information than its competitor.

Check your understanding

What was the ex ante forecast in the example?

Answer: 560 orders.

What was the ex post calculation using actual temperature?

Answer: 620.

Which error reflects the original forecast's operational performance?

Answer: The error of 75, not 15.

Related terms

Forecast horizon
Data leakage
Forecast origin
ex post forecast
real-time data vintage
rolling-origin evaluation
predictor uncertainty

Learn next

Create an availability timeline for every predictor in a forecasting pipeline. Then reconstruct one historical forecast using only information released before its issue time.

Sources

Hyndman-hosted Glossary of Forecasting Terms, Ex ante and Ex post forecast: https://robjhyndman.com/mwh3/FG4.pdf Hyndman and Athanasopoulos, Forecasting with regression, ex-ante versus ex-post forecasts: https://otexts.com/fpp3/forecasting-regression.html

Previous: EWMA chart Next: Exact algorithm