Symmetric mean absolute percentage error

wessel

Dear All,

I think it is good to include the SMAPE measure into the Rapid Miner regression performance operator.
http://en.wikipedia.org/wiki/SMAPE

This measure is more robust then correlation.
Allows to compare regression results from different data sets.
It is symmetrical. (Well not really, but better then MAPE)

Best regards,

Wessel


edit:
More detailed discussion:
http://www.buseco.monash.edu.au/ebs/pubs/wpapers/2005/wp13-05.pdf

I'm not really sure what error measure is best.
But it is clear that a measure is needed which is somewhat independent on the data set.
Like a measure that ranges from 0 to 1. Instead of -infinity to +infinity.

second edit:
What constitutes a good error measure? First, a good error measure is clearly interpretable and summarizes the cost related to the error. Second, a good error measure is robust to outliers. Third, a good error measure is stable if only a few data points are used. Finally, a good error measure is unaffected by units of measurement.