Is MAE sensitive to the error sign?

No; both directions contribute through absolute magnitude.

Does MAE penalize large errors?

Yes, linearly. It gives them less relative emphasis than squared error, but they still contribute to the average.

AI Glossary / Statistics and probability / Intermediate

Mean Absolute Error

Definition

Mean Absolute Error, or MAE, is the arithmetic mean of the absolute differences between predictions and their corresponding observed or reference values. For n cases, MAE = sum |y[i]-y-hat[i]| divided by n. It is nonnegative, equals zero for perfect agreement on the evaluated cases, and is expressed in the same units as the outcome.

What it means

MAE describes the average size of a prediction error without allowing overprediction and underprediction to cancel. If processing-time predictions have an MAE of three minutes, their absolute errors average three minutes over the evaluated cases.
This does not mean every prediction is within three minutes, nor that the next prediction will be three minutes wrong. The score summarizes a collection of errors and should be accompanied by information about their distribution when exceptional errors matter.

How it works

Match each prediction with the correct reference value. For a forecast, also match the prediction horizon and the information available when the forecast was issued. Comparing an updated prediction with an earlier target can accidentally evaluate a different task.
Subtract prediction from reference, or reference from prediction, and take the absolute value. Either sign convention gives the same MAE. Average the resulting nonnegative errors over the intended evaluation set.
Because errors are not squared, each additional unit of absolute error contributes the same amount to the total. A ten-unit error contributes ten times as much as a one-unit error. Under squared error it would contribute one hundred times as much.
This linear contribution makes MAE less dominated by large residuals than squared-error metrics, but not immune to them. A sufficiently large error can still raise the mean substantially. Report tail behavior separately if rare failures are especially consequential.
MAE treats overprediction and underprediction symmetrically. If shortages and surpluses have different costs, equal absolute distances may not have equal business consequences. Consider an appropriately asymmetric loss or an explicit cost calculation alongside MAE.
When choosing one constant prediction to minimize absolute error, any median of the outcomes is optimal. More generally, minimizing expected absolute error targets a conditional median when the relevant expectation exists. It does not generally target the conditional mean.
Evaluate on data that were not used to select the fitted model or tune its settings. A low training MAE can coexist with poor new-case predictions. For time series, use an evaluation design that respects time order and the forecast horizon.
For weighted evaluation, divide the weighted sum of absolute errors by the sum of weights. State whether weights represent frequency, exposure, or decision importance. Equal weighting of cases and equal weighting of groups answer different questions.
MAE depends on units and outcome scale. A score of five cannot be compared meaningfully across minutes, currency amounts, and temperatures without further context. Even within one unit, datasets with different target ranges or task difficulty can require a suitable baseline or scaled measure.
Finally, check the reference values. If labels or measurements are noisy, MAE describes agreement with those references, not necessarily error relative to an unobserved truth. Data quality and the evaluation population remain part of the interpretation.

Three practical examples

A scheduling model reports average absolute prediction error in minutes for a defined set of jobs.
A demand forecaster compares models on the same products, periods, and forecast horizons.
An AI measurement system is evaluated against validated reference measurements, with MAE reported in the measurement's units.

Think of it this way

Think of measuring the gap between each prediction and its target with a ruler, then averaging those gap lengths. The ruler ignores whether the prediction landed above or below the target. It does not reveal whether the errors consistently lean in one direction.

Worked example

An original example has observed values 10, 20, 30, and 40, with predictions 12, 16, 35, and 37. Using prediction minus observation, the signed errors are 2, -4, 5, and -3.
Their signed average is zero, because the positive and negative errors cancel. The absolute errors are 2, 4, 5, and 3, giving MAE = (2+4+5+3)/4 = 3.5. A zero signed average therefore does not imply accurate individual predictions.
For a separate constant-prediction illustration, consider outcomes 1, 2, and 10. Predicting their median, 2, produces absolute errors 1, 0, and 8, so MAE = 3.
Predicting their mean, 13/3, produces absolute errors 10/3, 7/3, and 17/3. Their average is 34/9, approximately 3.7778. The median does better under absolute error in this example; the mean instead minimizes squared error.
These calculations compare defined losses on the same observations. They do not establish which loss best represents a particular business decision.

Illustrative case study

In a fictional case, Willow Dispatch evaluates an arrival-time model using MAE. Two models have similar average errors, but one occasionally makes much larger late predictions.
The team keeps MAE as an average-distance summary and adds error quantiles and separate overprediction and underprediction checks. It then reviews how those errors affect dispatch decisions. The case illustrates the need to connect a metric to operations; it does not claim a measured improvement or a universal preference for one model.

Common misconceptions and mistakes

Do not confuse MAE with mean absolute deviation around a dataset's center. MAE compares each prediction with its own reference value.
Do not report MAE as a percentage unless a separate normalization has been defined. Do not assume it measures bias, because absolute values remove direction. A lower aggregate MAE can still conceal worse performance for an important subgroup.

Business use

MAE is useful when an average error in familiar units is easy to communicate and each unit of error has roughly similar importance. It supports clear model comparisons on a common evaluation task. Pair it with a baseline, uncertainty assessment, and decision-specific error checks so a small numerical gain is interpreted in context.

Check your understanding

How can the worked example have zero mean signed error and MAE 3.5?

Answer: Overprediction and underprediction cancel in the signed mean but remain positive distances in MAE.

Related terms

Learn next

Next, compare MAE with squared-error and quantile losses on the same predictions. Then design a held-out evaluation that matches the intended population, forecast horizon, and error costs.

Forecast horizon

Sources

Google Machine Learning Glossary, Mean Absolute Error: https://developers.google.com/machine-learning/glossary Scikit-learn, mean_absolute_error: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_absolute_error.html Hyndman and Athanasopoulos, Forecasting: Principles and Practice, forecast accuracy: https://otexts.com/fpp3/toolbox.html

Previous: Mean absolute deviation (m.a.d.) Next: Mean absolute percentage error