I read it on many sites, that one of the main "disadvantage" of MAPE is that it penalizes overpredictions, hence it prefers models that are under-predicting.

The main argument is that if we fix the forecasting value (e.g. 100) and the actual values are 50 or 150, then the MAPE's are 1 and 1/3.

For me it seems a bit silly, that we "fix" the forecasting value, but rather fix the actual value (I mean it is an actual value) and see how different forecasts behave. So actual value is 100, forecasts are 50 and 150. In this case, the MAPE's are both 0.5.

The other argument is that it does not have an upper bound. I dont think that this matters unless it's a rare type of data/model, but not something like a regular house repair prediction.

Am I missing something or MAPE is really not that bad in the aspect of punishing overpredictions?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.