1

The symmetric mean absolute percent error (SMAPE) is a symmetrized version of percent error with the formula:

$$\frac{200\%}{n}\sum_i\frac{|x_i - y_i|}{|x_i| + |y_i|}$$

SMAPE is symmetric: interchanging the $x$'s and $y$'s doesn't change the value. However, there are applications where it would be nice if an interchange of $x$ and $y$ changed the sign of the difference measure without changing the magnitude (ie if the measure was anti-symmetric).

For example, say that the value of an asset falls from $1 \rightarrow .95$, then recovers $.95 \rightarrow 1$. If you take the SMAPE of the price change events you get $5.13\%$. Instead I'd like to find an approach that would give an an answer of $0\%$ in this case.

Edit:

5 minutes after I wrote this question I thought of a simple answer (posted below). It would still be nice to know if there's an anti-symmetric percent difference measure that has actually been used before in the literature.

tel
  • 235
  • 2
  • 7
  • 1
    You are asking for a multiple of the "Relative Percent Difference." See https://stats.stackexchange.com/questions/86708 for a generalization. – whuber May 30 '19 at 19:45
  • @whuber Thanks. I think those are indeed the magic keywords – tel May 30 '19 at 20:23

1 Answers1

2

I suppose one obvious way of getting an anti-SMAPE would be to just use the plain, non-absolute, difference in the numerator:

$$\textrm{ASMAPE} = \frac{200\%}{n}\sum_i\frac{x_i - y_i}{|x_i| + |y_i|}$$

tel
  • 235
  • 2
  • 7