1
$\begingroup$

I am currently running an iterative process(for loop) to determine best ARIMA model for monthly sales data according to smallest AIC and MAPE. Box-Jenkins methodology clearly states to choose the order of the ARIMA components according to smallest AIC and validate according to out of sample accuracy measures, such as MAPE.

My question is:

Is over-fitting a risk if I choose my models according to smallest MAPE instead of AIC?Are there any risks/disadvantages?

$\endgroup$
  • 2
    $\begingroup$ Box-Jenkins methodology clearly states to choose the order of the ARIMA components according to smallest AIC does not seem to be a common interpretation of the Box-Jenkins methodology. Determining lag orders by ACF and PACF is a more common interpretation. $\endgroup$ – Richard Hardy Jan 23 at 19:08
  • $\begingroup$ You're right, thank you for the notice.The school subject were I learned about the methodology had mentioned the terms multiple times together. I guess the AIC came later in the method. $\endgroup$ – Andrei Catana Jan 24 at 9:45
2
$\begingroup$

Following @Richard Hardy's excellent comment ...and a paraphrase of @Adamo's advice here Interrupted Time Series Analysis - ARIMAX for High Frequency Biological Data?

"The correlograms i.e. the acf and the pacf should be calculated from residuals using a model that controls for intervention administration, otherwise the intervention effects are taken to be Gaussian noise, underestimating the actual autoregressive effect."

In other words if there is latent deterministic structure the acf and pacf can be quite misleading thus one needs to examine data that is conditioned upon possible pulses, level/step shifts, seasonal pulses and deterministic time trends.

Mindless fitting of all possible models can and often will produce redundancies and/or unnecessary differences and power transformations and over-populated structure. KISS (parsimony) should be the objective . Care should be taken to ensure that parameter constancy is validated over time and that model error variance is not flagrantly heterogeneous over time.

https://autobox.com/pdfs/ARIMA%20FLOW%20CHART.pdf suggests a script/paradigm that can often be useful.

$\endgroup$
  • $\begingroup$ Thank you, I remember my professor laying emphasis on the principle of parsimony. $\endgroup$ – Andrei Catana Jan 24 at 9:45
  • 1
    $\begingroup$ Einstein: “Everything should be as simple as it can be, but not simpler” $\endgroup$ – IrishStat Jan 24 at 9:49

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.