Can absolute error be negative?

No. Taking the absolute value makes it nonnegative. Zero means the prediction and reference match for that case.

Is absolute error the same as MAE?

No. Absolute error describes one discrepancy; MAE averages absolute errors across cases, subject to any stated weighting.

Does lower MAE always mean the better business choice?

Not necessarily. The evaluation data, important subgroups, extreme errors, and different costs of over- and underestimation can affect the decision.

AI Glossary / Graphs and symbolic AI / Beginner

Absolute error

Definition

Absolute error is the size of the difference between a predicted or estimated numerical value and its reference value, without regard to direction. It is calculated as |prediction − reference|. A prediction five units too high and one five units too low both have absolute error five. The result retains the original measurement unit.

What it means

Absolute error answers a simple question: how far away was the estimate? It does not indicate whether the model overestimated or underestimated, because taking the absolute value removes the sign.

For multiple predictions, mean absolute error, or MAE, averages the individual absolute errors. The individual error and its dataset average are related but different quantities. Reporting one large error is not the same as reporting the average across all cases.

The reference value also matters. It might be a recorded outcome, a trusted measurement, or an agreed target. If that reference is inaccurate or measured under different conditions, the calculated error may not fairly describe the prediction's quality.

How it works

Match each prediction to the correct reference for the same case, unit, and time period. Subtract one from the other and take the absolute value. The result is nonnegative and is zero when the two values match exactly.

To calculate unweighted MAE, add all individual absolute errors and divide by the number of cases. Weighted versions give some cases more influence, so the weighting should be stated. Multiple outputs measured in different units should not be averaged indiscriminately without a meaningful evaluation design.

Unlike squared error, absolute error grows linearly with the size of the discrepancy. Doubling an error doubles its absolute-error contribution rather than quadrupling it. This can be useful for interpretation, but it does not establish that the business cost of an error grows linearly or that overestimation and underestimation are equally costly.

Three practical examples

1. A demand forecast predicts 120 items and actual demand is 108. The absolute error is 12 items, providing a direct measure of the forecast's numerical miss.

2. An estimated journey time is 35 minutes and the observed time is 42 minutes. The absolute error is seven minutes, although the direction also matters to someone planning an arrival.

3. A sensor-estimation model is compared with a reference measurement in the same units. Absolute errors help reviewers see the size of discrepancies while separate checks examine systematic over- or underestimation.

Think of it this way

Think of measuring the distance between a dart and a target point along a marked line. The distance tells you how far the dart missed, regardless of whether it landed to the left or right.

The analogy is limited. Many real outcomes have different costs on opposite sides of the target, and a single numerical distance can hide that distinction. It also assumes that the target position is reliable. A poorly measured reference can make an estimate appear wrong for reasons unrelated to the model.

Worked example

Consider four original toy forecasts, with predicted quantities 18, 27, 35, and 50, and observed quantities 20, 24, 40, and 46. Using prediction minus observation, the signed errors are −2, 3, −5, and 4.

Their absolute errors are 2, 3, 5, and 4 units. The total absolute error is 14, so MAE is 14 / 4 = 3.5 units.

The mean signed error is (−2 + 3 − 5 + 4) / 4 = 0. A zero mean signed error does not mean every forecast was correct; positive and negative errors canceled. The positive MAE reveals the average size of the misses.

The same 3.5-unit MAE could arise from a different pattern of errors. To understand reliability, inspect the distribution and individual consequential cases rather than assuming that every prediction is about 3.5 units away from its outcome.

Illustrative case study

Fictional case: Birch Catering evaluates a model predicting daily meal demand. The dashboard reports a low average signed error, and managers initially assume the forecasts are consistently close to actual demand.

An analyst calculates absolute errors and finds that substantial overestimates and underestimates have been canceling. The team begins reviewing MAE alongside the direction of errors and the operational consequences of shortages and surplus food.

This invented case claims no numerical improvement. It illustrates why error magnitude and error direction answer different questions. A useful evaluation should reflect the decisions the forecast supports, including whether one type of mistake is more costly or difficult to recover from than another.

Common misconceptions and mistakes

Absolute error is not percentage error. An error of ten items has the same absolute size whether actual demand was twenty or two thousand, but its relative size differs greatly.

MAE is also not the maximum error, nor a promise that future errors will stay below the reported average. Comparisons require compatible units and evaluation data. A lower MAE on easier cases does not necessarily indicate a better model. Finally, an observed value of zero is valid for absolute error; no division by the reference is required.

Business use

Businesses can use absolute error when the original unit is meaningful, such as minutes, items, or measured quantities. It makes individual prediction misses easy to explain and supports straightforward aggregation across comparable cases.

Choose evaluation periods and samples that reflect the intended use. Report relevant groups and unusual failures alongside the average, and retain signed errors where direction matters. If business costs are asymmetric or nonlinear, assess those costs separately instead of treating absolute error as a complete financial or operational objective.

Check your understanding

Predictions have signed errors of +8 and −8 units. What are the mean signed error and MAE?

Answer: The mean signed error is zero, while MAE is eight units. Cancellation removes direction from the first average but does not remove the magnitude of the two misses.

Related terms

Mean Absolute Error
Residual
Root Mean Squared Error
Forecasting
signed error
squared error
percentage error
regression

Learn next

Next, compare absolute and squared error on the same small set of predictions. Then identify whether overestimating and underestimating have the same practical consequences for the task you are evaluating.

Sources

Scikit-learn, mean_absolute_error documentation: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_absolute_error.html The primary metric documentation was read. This entry distinguishes individual absolute error from its mean and uses original calculations and business examples. The supplied graphs-and-symbolic-AI category is retained, although the metric applies broadly to numerical estimation.

Previous: Ablation Next: Absolute error loss