Mean absolute error (MAE) is a measure of forecast accuracy. It is the average size of the differences between a set of forecasts and the values that were later observed, taken without regard to whether each forecast was too high or too low. The result is in the same units as the thing being forecast: degrees for a temperature forecast, dollars for an earnings estimate, percentage points for an inflation projection.
MAE is one of the standard scores used to compare forecasting models in meteorology, economics, finance and machine learning. A lower value means the forecasts were closer to the outcomes on average. It is usually reported alongside the root mean squared error, which penalizes large misses more heavily, and the mean error, which shows whether the forecasts lean in one direction.
Formula and example
The standard calculation is:
- is the forecast for case .
- is the actual value observed for case .
- is the number of forecasts.
- The vertical bars denote the absolute value, which drops the sign of each error.
Suppose a model forecasts a city's daily high temperature for five days as 70, 72, 68, 75 and 71 degrees, and the observed highs are 72, 71, 65, 75 and 74. The errors are -2, 1, 3, 0 and -3. Their absolute values sum to , so the MAE is degrees.
The mean error of the same forecasts is degrees. The misses nearly cancel, so the model shows almost no bias even though its typical miss is close to two degrees. This is why the absolute value is taken.
Comparison with other error measures
The root mean squared error squares each error before averaging and then takes the square root:
For the example, the squared errors sum to , and the RMSE is degrees. RMSE is never smaller than MAE, and the gap between the two widens when a few errors are much larger than the rest. MAE treats a six-degree miss as exactly twice as bad as a three-degree miss. RMSE treats it as four times as bad before the root is taken.
The choice between them depends on the cost of errors. Where a large miss is disproportionately harmful, RMSE is the better guide. Where cost rises in proportion to the size of the miss, MAE is more faithful and is less distorted by a single outlying case. In statistical terms, the forecast that minimizes expected absolute error is the median of the possible outcomes, while the forecast that minimizes squared error is the mean.
The mean absolute percentage error divides each error by the actual value, which allows comparison across series of different scale but fails when actual values are near zero. For forecasts stated as probabilities, the Brier score is used instead, and forecast calibration is assessed separately.
Use in weather and event markets
Meteorological agencies publish MAE figures for temperature, wind and rainfall forecasts at each lead time, and the score is a main yardstick for comparing numerical weather prediction models. Statistical post-processing such as model output statistics is judged by how far it lowers the MAE of raw model output at a given station. Temperature MAE grows with lead time, from around two degrees Fahrenheit a day ahead to several degrees a week ahead in typical conditions.
These figures matter to traders of any weather derivative and of an event contract that pays according to an observed temperature. A contract on whether a city's high will fall within a two-degree band cannot be priced with confidence from a forecast whose typical miss is also about two degrees. Participants in a prediction market therefore use the historical MAE of each model, by station and season, to turn a point forecast into a spread of probabilities. An ensemble forecast serves the same purpose from the model side.
Use in finance and machine learning
In finance, MAE is used to grade analysts' earnings estimates against reported results, to test models for volatility forecasting, and to measure cash forecast accuracy in corporate treasury. In machine learning it serves both as a score for a fitted model and as a loss function during training, where it makes a regression less sensitive to extreme observations than squared-error loss.
Limitations
MAE compresses a whole distribution of errors into one number. It hides bias, since it ignores sign, and it hides the occasional very large miss. It depends on the units and scale of the data, so values for different variables cannot be compared directly. A score computed on the data used to build a model understates the error to be expected on new data, a symptom of overfitting, and should be checked on a separate sample.