Symmetric Mean Absolute Percentage Error

Computes the symmetric mean absolute percentage error (SMAPE) between predicted and reference values.

Usage

smape(x, ...)

# S3 method for class 'lm'
smape(x, ...)

# Default S3 method
smape(x, ref, na.rm = FALSE, ...)

Arguments

x

an object. Methods are available for numeric vectors and model objects (e.g. lm).

...

additional arguments passed to methods

ref

numeric vector of reference (true) values

na.rm

logical; whether to remove missing and undefined terms

Value

a numeric scalar containing the SMAPE

Details

The SMAPE is defined as: $$ \frac{1}{n} \sum \frac{2 |ref - x|}{|x| + |ref|} $$

Values are bounded between 0 and 2. Note that this is a ratio, not a percentage: the factor 100 of the original definition is not applied, which is the form for which the bound of 2 holds.

A term is undefined when x and ref are both zero, and is set to NA. With the default na.rm = FALSE a single such pair therefore makes the whole result NA; with na.rm = TRUE those terms are dropped along with genuinely missing ones, so the mean is taken over fewer than length(x) terms.

See also

Other model.metrics: auc(), averagePrecision(), brierScore(), logLoss(), mae(), mape(), mse(), nmae(), nmse(), rmse()

Examples

x <- c(2.5, 3.0, 2.8)
ref <- c(3.0, 2.5, 3.0)

smape(x, ref)
#> [1] 0.1442006

# with linear model
fit <- lm(mpg ~ hp, data = mtcars)
smape(fit)
#> [1] 0.1550795