A forecast that was ten orders too high one day and ten too low the next has an average error of zero. That zero says the misses cancelled out, not that the forecasts were right. Reviewing outlet forecasts needs two measures, one for direction and one for size, and a clear statement of what each row actually compares.
This guide reviews forecasts already issued for completed service periods at more than one outlet. It does not build forecasts, plan capacity, approve service or revenue, rank staff, or replace a wider financial scorecard. The example, FR-025, is original fiction, and its small sample supports no statistical conclusion.
Fix what is being forecast before looking at any result
Every row must use the same definition. In FR-025 the measure is completed dine-in orders, each order counted once, for one outlet and one service period in one fixed timezone. Channel, period boundaries and the forecast lead policy are written down before review. Here the policy is day-ahead: each service's forecast is issued at 18:00 the previous day.
Record the forecast version, issue time and source for every target, and confirm it was frozen before the target service began. A revision after service starts is not the same day-ahead forecast. A legitimate forecast for a remaining interval needs its own target and horizon; it must not replace the original full-service forecast in this review. Actuals are recorded only once they are final and verified after service, with their own source and finalisation time.
Completed orders are what was recorded, not unconstrained demand. If an item ran out, a section closed, the weather turned or a channel changed, note it as context. Do not estimate the orders that might have happened and add them back.
Register every planned target first
Before any actual is known, list every outlet-service pair the review will cover and record an independent count of planned targets. In FR-025 that count is six: North and South, each for dine-in service from 08:00 to 20:00 IST on 4, 5 and 6 November 2026. Forecast versions F04, F05 and F06 cover the three dates, each actual is verified final by 21:00 on its target day, and the review cutoff is 23:00 on 6 November.
Those planned rows stay. A target with a missing actual is not dropped, and missing is not zero: its errors stay blank and every pooled figure is held back until it is resolved. A verified actual of zero is a valid known value. Predeclared exclusions, with reasons, go in a separate note; never remove a row because its result looks bad.
Signed error shows direction; absolute error shows size
For each row, the error is the actual minus the forecast. A positive error means more orders than forecast; a negative one means fewer. The absolute error drops the sign and shows how far off the forecast was. Averaging signed errors gives the mean error, a direction pattern. Averaging absolute errors gives the mean absolute error, or MAE, a size measure in the data's own unit, here orders per outlet-service.
Worked example: FR-025
The table lists the six frozen day-ahead forecasts and their final verified actuals, in completed dine-in orders.
| Outlet and date (fictional) | Forecast | Actual | Signed error | Absolute error |
|---|---|---|---|---|
| North, 4 Nov | 100 | 110 | +10 | 10 |
| North, 5 Nov | 110 | 100 | −10 | 10 |
| North, 6 Nov | 120 | 120 | 0 | 0 |
| South, 4 Nov | 50 | 40 | −10 | 10 |
| South, 5 Nov | 60 | 70 | +10 | 10 |
| South, 6 Nov | 70 | 80 | +10 | 10 |
Reading the outlet results
North's signed errors were +10, −10 and 0, so its mean error is 0. Its absolute errors were 10, 10 and 0, giving an MAE of 20 ÷ 3, about 6.67 orders per outlet-service. A mean error of zero does not mean North forecast perfectly; two of its three forecasts missed by ten orders.
South's signed errors were −10, +10 and +10, so its mean error is 10 ÷ 3, about 3.33: actuals tended to come in above forecast over these three days. Its MAE is 10. That is an observed pattern in three rows, not proof of systematic bias and not a cause.
Do not rank the outlets on 10 against 6.67 without context. They serve different volumes, and three fictional days carry no statistical weight. Comparing MAE across measures in different units is invalid, and even within one unit, volume, horizon and mix still matter. A measure such as mean absolute percentage error divides each error by its actual value: it is undefined at zero and can become extreme near zero. This worksheet uses no percentage accuracy score.
Pooled outlet-service MAE is not chain-daily MAE
Pooling all six rows gives a total forecast of 510 against an actual of 520. The signed total is +10, so the pooled mean error is 10 ÷ 6, about 1.67, and the pooled MAE is 50 ÷ 6, about 8.33 orders per outlet-service.
That is not the error of a chain-level daily forecast. Adding both outlets for each day gives forecasts of 150, 170 and 190 against actuals of 150, 170 and 200. The daily absolute errors are 0, 0 and 10, a mean of about 3.33 orders per chain-day. The gap comes from cancellation: on 4 and 5 November the outlets missed in opposite directions, and adding them together hides both misses. Label the level of every summary, and never quote one where the other is meant.
Comparing methods, corrections and follow-up
To compare two forecasting methods, compare forecasts produced for the same targets using the same information cutoff. A fair review of past unseen outcomes can inform the next method choice. Refitting a method with those outcomes and presenting revised historical predictions as the original forecasts cannot. Check a chosen method on later unseen periods before claiming its advantage persists. A transcription correction or actuals restatement gets its reason, evidence, reviewer and date while preserving the original; a revised forecast after the information cutoff is a separate horizon, not a repaired old prediction. Each finding gets an owner and a follow-up action.
Where the measures come from
Forecasting: Principles and Practice, third edition, explains that accuracy should be judged on genuine forecasts rather than the fit to data used to build a model, defines a forecast error as the observed value minus its forecast, describes the mean absolute error, and notes that scale-dependent measures cannot compare series in different units while mean absolute percentage error is undefined when an actual is zero. Only those points are used here. Its code, examples, forecasting methods, test-set conventions, other measures and model recommendations are not adopted, and no Keystonne feature is implied. FR-025 and its workflow are original fiction.
Put it into practice
Your action checklist
Your progress saves in this browser. No sign-up needed.
0 of 10 complete
From reading to doing
Outlet Forecast Review
Open Outlet Forecast Review for one review case of up to 12 unique outlet-service targets, registered before any outcome is known. Enter the case reference, metric definition, unit, source, forecast horizon policy, one fixed timezone offset and review cutoff, an independent planned target count, and confirmations that the planned frame is complete and every row shares a comparable basis. Each row holds a unique target ID, outlet, full typed target start and end date-times, a forecast issue time before the start, a non-negative forecast (fractions allowed), a non-negative whole-number actual, an actual finalisation time no earlier than the end and no later than the review cutoff, and forecast version, forecast source and actual source references. Enter the fixed lead and target duration in whole minutes. The example uses a lead of 840 minutes and a duration of 720 minutes; row times must match those scope values. Times use whole minutes and end must follow start; duplicate target IDs or outlet-and-period pairs are flagged. Signed error (actual minus forecast) and absolute error show only on valid rows. Pooled totals, mean error and MAE appear only when every planned row is valid and complete and the row count matches the independent target count; a missing actual withholds them rather than dropping the row. Zero planned rows shows No target services, with no averages or totals. Rows may sit in any order with gaps, and wholly blank rows are ignored. Per-outlet readouts are checked manually, as in the example. There is no percentage accuracy score, MAPE or automatic tolerance. Manual sections hold predeclared context and exclusions with retained reasons, restatements that preserve originals, and review actions with owners and follow-up.
Open the Outlet Forecast Review tab in the launch workbook. Keep a copy for your own business.
Download the free workbook See what’s inside the toolkit →Sources & further reading
- Evaluating point forecast accuracy (Forecasting: Principles and Practice, 3rd edition, OTexts) ↗
Retrieved 8 November 2026. Used only for the general points that accuracy is judged on genuine forecasts rather than in-sample fit, that a forecast error is the actual minus the forecast, that MAE summarises error size, that scale-dependent measures cannot compare series in different units, and that mean absolute percentage error is undefined when an actual is zero. No code, datasets, forecasting algorithm, test-set sizing rule, MASE, RMSE or model recommendation is adopted. FR-025, its figures and workflow are original fiction.
Published by Keystonne Grow. Product links connect to the Keystonne ecosystem; check the linked product or service for its current scope.