4
$\begingroup$

I am looking at a demand forecasting tool which is considered an industry standard in the retail domain, and it offers several forecasting methods, most of which are some variation of Exponential Smoothing. It doesn't offer (or even mention) ARIMA at all, which surprises me, given how important ARIMA seems to be, based on the literature.

Is there any reason why they would go with Exponential Smoothing over ARIMA? Is Exponential Smoothing better for demand forecasting than ARIMA? Or is it the case that most useful ARIMA models can be re-cast as Exponential Smoothing models?


Since I first posted this question, I came across the following paper "Why the Damped Trend works" by Gardner and McKenzie in which the authors argue (also see the references within) to the effect that exponential smoothing with damped trend can “reasonably claim to be a benchmark forecasting method for all others to beat.”

$\endgroup$
  • 2
    $\begingroup$ I am not familiar with the industry, but I suppose ES is much easier to explain to your statistically illiterate superior. On a more methodological note, you can see ES as a way of constructing a nonparametric smoothing technique (I.e., you can define the exponential weights via a kernel). If you use that as a motivation, you could make the argument that the implied assumptions on your time series for your prediction to be robust are weaker than in the ARIMA case. $\endgroup$ – Jeremias K Dec 11 '17 at 22:15
2
$\begingroup$

Exponential smoothing models are in general a subset of ARIMA models . When I say ARIMA models I am including the possibility of including trends, level shifts ,seasonal pulses and pulses in the equation. This is also known as a subset of ARMAX models. ARIMA models are more general thus requiring some logic in forming a final useful model. Demand planning software is deficient in this regard even from the so-called leaders.

In terms of why one would go with es over ARIMA it is quicker being more presumptive). Additionally poor ARIMA model identification sftware assuming no anomalies often leads to inadequate forecasts.

If you want to conduct a mini-experiment of es models and ARIMA models for a few time series , I would be glad to help but offline as this kind of task is beyond SE's mission.

$\endgroup$
1
$\begingroup$

https://www.otexts.org/fpp2/arima-ets.html

Ultimately, one would need to perform time series cross-validation to determine which will perform better in any individual case.

$\endgroup$

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.