MAE

MAE. Mean Absolute Error (MAE) is the average per-meal absolute prediction error in kilocalories, before normalization by ground-truth value. It complements MAPE by preserving the absolute size of the error.

What is MAE?

MAE is the average absolute error in kilocalories. For meal i with ground-truth value yᵢ and predicted value ŷᵢ, the per-meal absolute error is |yᵢ − ŷᵢ|. MAE is the mean over n meals.

Unlike MAPE, MAE preserves the absolute size of the error. A 5% MAPE on a 200-kcal meal is a 10-kcal MAE; on an 800-kcal meal it is a 40-kcal MAE. The asymmetry can matter for specific use-case judgments.

When MAE is the right metric

MAE is informative when the question is “how big is the typical per-meal error in kilocalories I would experience?” rather than “how accurate is the tracker as a percentage on a typical meal?”

Methodology v3.2 reports MAE alongside MAPE in supplementary tables. The ranking of apps by MAE is roughly the same as the ranking by MAPE for our 50-meal battery, because the meal-size distribution in the battery is roughly even.

How it differs from MAD

MAE and MAD (mean absolute deviation) are sometimes used interchangeably in less-careful sources. In v3.2:

• MAE is the mean absolute prediction error against ground truth.
• MAD is the mean absolute deviation around the median of the error set, used as a dispersion measure.

The two answer different questions. MAE: how big is the typical error. MAD: how concentrated is the error distribution around its center.

For full discussion, see our metric-comparison article.