With high variabilty data, how do you choose between ETS and ARIMA?
Accuracy measures are often interpretive making it hard to choose which model is the best. The rule of thumb is to have lowest numbers possible associated with percent of error. With that said, numbers sometimes flip between models (ETS and ARIMA). Looking for opinions on the Root Mean Squared Error (RMSE) - Mean Absolute Error (AMA); the Mean Percentage Error (MPE), the Mean Absolute Percentage Error (MAPE), and the Mean Absolute Scaled Error (MASE). The various errors can be better in between models. What is your process to decide which model you choose? (this is with consideration that historical data aligns with trends, seasonality and moving averages).
Any thoughts?
Answers
-
Hello Ray,
My name is Spencer and I am one of the moderators for the TSIA Community Exchange. This is an amazing question! You have provided an excellent amount of detail on a topic that is rife with complexity and nuance. I am sure that a dialogue on this topic is going to prove fruitful. I am going reach out to a couple of my contacts and see if they'd like to hop into this thread and provide their perspective. In the meantime, here are a TSIA artifacts that you may find relevant:
RM in a Rapidly Virtualizing PS World
0 -
@Iurii Kim and @Anthony Medeiros , I wonder if this discussion might be up your alley?
0
