Can a decrease in mean absolute percentage error (MAPE) be correlated with an increase in standard deviation of the error? Is that counter-intuitive? What about MAPE vs. mean absolute error (MAE)? I'm wondering if those metrics should follow a trend, or not necessarily.


The general concern is as follows: when performing cross-validation, by establishing a scoring criterion that combines average (e.g., MAE, MSE) and spread (e.g., SD) of the errors doesn't necessarily lead to minimizing MAPE. In other words, one shouldn't expect to reduce MAPE by minimizing those other metrics. Does this sound reasonable?


Note that the MAPE goes down as actuals go up - and the standard deviation doesn't. So for a given time series of errors (with potentially increasing SD), we could simply have a time series of actuals with a positive trend, and once the positive trend is strong enough, the MAPE will start going down.

nn <- 100
error <- rnorm(nn,0,seq(10,15,length.out=nn))
actuals <- seq(20,50,length.out=nn)

cumulative.mape <- cumsum(abs(error)/actuals)/(1:nn)
cumulative.sd <- sapply(1:nn,function(xx)sd(error[1:xx]))

opar <- par(mfrow=c(1,2))
    plot(cumulative.mape[-(1:10)],type="l",main="Cumulative MAPE",ylab="",xlab="")
    plot(cumulative.sd[-(1:10)],type="l",main="Rolling Error SD",ylab="",xlab="")


So the issue is that the MAPE depends on the error and the actuals, whereas the SD of the errors don't depend on the actuals any more (beyond the actuals influencing the errors themselves, of course). Thus, this should typically not happen for the SD and MAE, since the MAE again only depends on the errors, not the actuals.

EDIT: In general, different error measures move somewhat in tandem - but not perfectly so. Minimizing different error types is the same as optimizing different loss functions - and the minimizer for one loss function is typically not the minimizer of a different loss function.

For an extreme example, minimizing the MAE will pull you towards the median of the future distribution, while minimizing the MSE will pull you towards its expectation. If the future distribution is asymmetric, these will be different, so minimizing the MAE will yield biased predictions. I just discussed this yesterday.

So: no, minimizing one error measure will not necessarily minimize a different one.

I regularly read the International Journal of Forecasting, and accepted best practice there is to report multiple error measures, and sometimes, yes, they imply that different methods are "best". Which authors and readers take in stride. I'd say that point forecasts are not overly helpful, anyway, and that you should always aim at full predictive densities.

(Incidentally, I can't recall ever having seen the SD of the errors reported in the IJF, and I don't really see the point of it as an error measure. An error time series can be badly biased and constant over time, with a zero SD - what's good about that?)

| cite | improve this answer | |
  • $\begingroup$ Thanks for that. So it seems that one can't generalize trends of MAPE to those of SD. So it wouldn't make sense to minimize one and expect the other to also decrease necessarily, right? I updated the question with my general concern. $\endgroup$ – Bruno May 30 '17 at 16:34
  • $\begingroup$ You don't actually need a trend, though; consider the MAPE of a series with $\mu = \sigma = 1$ vs the MAPE of a series with \$mu = 100, \sigma = 2$, although this doesn't affect your fundamental point. $\endgroup$ – jbowman May 30 '17 at 16:46
  • $\begingroup$ I edited my answer in reply to your edit. Hope it helps! $\endgroup$ – Stephan Kolassa May 30 '17 at 17:03
  • $\begingroup$ Thanks a lot, Stephan! The motivation for minimizing SD in addition to MSE or MAE (it seems MSE is preferable, right) is to not only attempt to make the error distribution zero-centered, but also "thin" (i.e., small spread). Any comments on that? $\endgroup$ – Bruno May 30 '17 at 17:54
  • 1
    $\begingroup$ Yes, they can. wMAPE is the sum of the absolute errors over the sum of the actuals (see here), which works just as well for non-time series. mase divides the MAD by the MAD achieved in-sample by a benchmark method - typically we use the random walk, where this error is $Y_t-Y_{t-1}$, but any other reasonable benchmark works as well. See here, or other threads tagged mase. $\endgroup$ – Stephan Kolassa May 31 '17 at 12:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.