I have recently been refreshing my forecasting knowledge while working on some monthly forecasts at work and reading Rob Hyndman's book but the one place I am struggling is when to use an exponential smoothing model vs an ARIMA model. Is there a rule of thumb where you should use one methodology vs another?

Also, since you can't use AIC to compare the two do you just have to go by RMSE, MAE, etc?

Currently I am just building a few of each and comparing the error measures but I wasn't sure if there was a better approach to take.

  • $\begingroup$ As I recall from Hyndman's book , a major point of smoothing techniques is to smooth the data. It doesn't consider noise or volatility of the noise. It can be used for predictions, but that doesn't seem to be the main point. $\endgroup$
    – meh
    Commented Jun 30, 2016 at 18:11
  • 3
    $\begingroup$ @aginensky, exponential smoothing is definitely a popular and effective forecasting technique. I would guess the main use of the exponential smoothing models is nothing else but forecasting. $\endgroup$ Commented Nov 13, 2016 at 17:03
  • $\begingroup$ That's correct, in fact until recently there was no such thing as an exponential smoothing model; exponential smoothing was only an algorithm for computing forecasts, nothing else. $\endgroup$
    – Chris Haug
    Commented Jan 13, 2017 at 17:50

2 Answers 2


Exponential Smoothing is in fact a subset of an ARIMA model. You don't want to assume a model, but rather build a customized model for the data. The ARIMA process let's you do that, but you need to also consider other items. You need to identify and adjust for outliers also. See more on Tsay's work with outliers here

  • 3
    $\begingroup$ In a broad sense, exponential smoothing is not a subset of ARIMA models, although linear exponential smoothing models indeed are. See Hyndman & Athanasopoulos "Forecasting: Principles and Practice" Section 8.10. $\endgroup$ Commented Jul 9, 2016 at 12:08
  • $\begingroup$ Yes, you are correct. It's all true that there are ARIMA models with no ETS counterpart. readbag.com/robjhyndman-research-rtimeseries-handout Would have an example dataset you can point me to that would be a good benchmark for this? $\endgroup$
    – Tom Reilly
    Commented Jul 9, 2016 at 16:43
  • $\begingroup$ I don't have a good data set for benchmarking, no. $\endgroup$ Commented Jul 9, 2016 at 16:45
  • $\begingroup$ I should add that Autobox(a software I am a part of ) doesn't restrict the coefficient <1 so for Autobox it does mimic non-linear properties. ETS also ignores 1) Pulses, Level Shifts, Seasonal Pulses and one and only 1 trend ; 2) constancy of error variance ; 3) constancy of parameters over time. $\endgroup$
    – Tom Reilly
    Commented Jul 9, 2016 at 16:49

I've performed a fairly extensive testing of ARIMA, Holt winters and others and tabulated the results here.

It's notable that ARIMA(3,0,0) does pretty well, as does ARIMA(2,0,1) across a pretty wide range of time-series, but of course you should see what works for your problem.


Your Answer

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

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