Using MAPE for Autoregression

[Taliah]

Hello, Do you know if MAPE can be used to asses an autoreg model. Using proc Autoreg, it is listed in the results, but we understand that some statisticians do not like to use it (maybe we misunderstood). If it is not a good way to asses the model, which of the other measures provides a good assessment? Thank you!

[sbxkoenk · Accepted solution]

Choosing such a goodness-of-fit / quality assessment metric is always a challenge. They all have their advantages and disadvantages and it also depends a lot on what you as a modeler want to achieve with the model : Explaining the past ... forecasting the future ... accuracy ... explain-ability and interpretability ... conciseness ... .

Personally, I prefer to use AICc (or BIC) because MAPE & RMSE only look at the error without taking into account the complexity of the model. RMSE/MAPE brings therefore more risk of overfitting, while AICc/BIC uses a penalty for each additional term in the model.

An "infinite" amount has already been written about this. You're sure to find good resources. I just hope they don't just make the choice harder.

[Taliah]

Thank you very much for all the details!