Root Mean Square Error (RMSE)

Root Mean Square Error (RMSE) is the square root of the mean of squared differences between estimated and reference values. Like MAE, it expresses accuracy in the units of the thing being measured — kilocalories, grams, milligrams. Unlike MAE, it punishes large errors disproportionately: a single 500-kcal estimation error contributes 250,000 squared kilocalorie-units to the sum, while ten 50-kcal errors contribute 25,000 in total. The squaring is the point.

The formula

RMSE = sqrt((1/n) × Σ (ê_i − e_i)²)

The sequence is: compute the signed error per observation, square each, average, take the square root. The final square root restores the original units — kcal, not kcal² — so a reported RMSE can be compared directly to the reference values.

RMSE vs MAE

Mathematically, RMSE ≥ MAE for any set of errors, with equality only when all errors have the same magnitude. The gap between them is a rough gauge of error heterogeneity: a method with RMSE near MAE produces errors of roughly consistent size, while a method with RMSE much larger than MAE produces a skewed distribution with a small number of large errors dominating. Both statistics should be reported when the application allows.

When RMSE is the right choice

RMSE is appropriate when large individual errors have disproportionately bad downstream consequences. Three examples from the dietary-assessment literature:

• Insulin dosing in type 1 diabetes. A person using pre-meal insulin dosed to a carbohydrate-count estimate faces glycaemic harm from under-dosing that is not linearly scaled — a 40-gram underestimate and a 4-gram underestimate differ by far more than a factor of 10 in blood-glucose consequence. Carbohydrate-estimation methods for this population are typically benchmarked by RMSE.
• Clinical enteral-nutrition support. Tube-fed patients receive calculated calorie targets; a 20 per cent miss on a single day is clinically meaningful.
• Inborn error of metabolism management. Phenylketonuria and similar conditions require tight daily protein bounds; RMSE captures the bounding violations better than MAE does.

When MAE is the right choice

For ordinary consumer calorie tracking, where the downstream consequence of a mis-estimated meal is a modest miss on a weekly average weight target, the cost function is roughly linear in error magnitude and MAE (or MAPE) is the more interpretable measure. A large majority of consumer-facing accuracy claims in this literature use MAE or MAPE rather than RMSE. An app quoting RMSE in consumer marketing without context is typically using it because RMSE is a slightly more flattering number to quote when the error distribution has few outliers (since RMSE = MAE in that edge case, there is nothing to lose).

Reporting standard

Any RMSE figure should carry the reference set, the sample size, and — ideally — the corresponding MAE alongside, so readers can judge the error distribution's shape. A 2019 methodological review in Public Health Nutrition on wearable-device calorie estimation argued that dietary-assessment papers should report RMSE, MAE, and signed bias together by default, treating any single-statistic report as insufficient.