What is mean absolute error?
Mean absolute error (MAE) is the average size of a model’s errors, ignoring whether each prediction was too high or too low. For each case, take the difference between the predicted and actual value, drop the sign, and average the results. If a forecast misses by 10, 15, and 5 units on three days, the MAE is 10 units. The result is in the units of the quantity being predicted, which makes MAE easy to explain.
MAE weighs every error in proportion to its size, so one miss of 30 counts the same as three misses of 10. Root mean square error (RMSE) squares errors before averaging and punishes large misses more. MAE suits problems where a big error costs no more than its size suggests.
Mean absolute error in practice
A managed service provider forecasts daily help desk tickets to set technician schedules. A typical day brings about 220 tickets, and one technician resolves about 25. The model’s MAE is 14 tickets, roughly half a technician’s day, against 31 for the old method of copying the same weekday from the previous week. The service desk manager could judge the gain without any statistics.
MAE cannot compare items on different scales, since an error of 14 tickets is serious at 220 a day and trivial at 2,000. Mean absolute percent error (MAPE) handles that comparison. A model tuned to minimize MAE also aims at the median outcome, so for a quantity that is zero on most days it can learn to predict zero.