3
$\begingroup$

I am working with the following monthly time series to build forecasting models: enter image description here

From this plot, it's quite tricky to identify if there is some kind of seasonality or not. When I use the ets function, it comes up with a ETS(M,N,N) which I believe suggests non-seasonal data. However, my auto-arima function comes up with a ARIMA(0,1,0)(1,1,0)[12] which suggests seasonality, right? Why do the algorithm disagree? Am I interpreting it right?

Also, would you recommend a way to identify any potential seasonality in my time series? Ideally, I would like to make a preliminary analysis where I can state whether my time series is most likely seasonal, or not. Thank you.

This is the data I'm using:

structure(c(198.5, 206.5, 220.25, 210, 201.5, 235, 248, 249, 
243, 236.25, 236.25, 260.75, 282, 271, 298.8, 351.5, 324.9, 307.8, 
320, 322.6, 334.7, 344.6, 366.3, 395.2, 423.3, 462.4, 435, 440.6, 
412.7, 437.3, 460.9, 453.5, 455.8, 431.4, 424.4, 461.3, 484.7, 
500, 473.7, 435.7, 426.5, 425.2, 440, 320, 349.5, 368.4, 365.7, 
314.7, 329.5, 354.6, 315.9), .Tsp = c(2012.75, 2016.91666666667, 
12), class = "ts")
|cite|improve this question
$\endgroup$
  • 1
    $\begingroup$ Can you edit your question to include your data? Best to put in the output of dput(mydata). $\endgroup$ – Stephan Kolassa Nov 21 '17 at 20:39
  • 1
    $\begingroup$ ETS won't identify data that needs to be differenced to be stationary as such; it assumes the underlying process is stationary and works from there. Try differencing the data, then running the differenced data through ETS and see what happens (also plot the differenced data, of course.) $\endgroup$ – jbowman Nov 21 '17 at 21:02
  • $\begingroup$ Hi, thanks for the hint. I tried your suggestion and made the data stationary before running ETS, now I get ETS(A,N,N). Still not considering it as seasonal. $\endgroup$ – Notna Nov 21 '17 at 21:09
  • 1
    $\begingroup$ @jbowman: actually, ETS does not assume stationarity. It will explicitly model trend and seasonality. One important aspect is that both are modeled locally, i.e., the trend can change, similarly to Holt-Winters. $\endgroup$ – Stephan Kolassa Nov 22 '17 at 7:40
  • $\begingroup$ @StephanKolassa - yes, you're right, I was just thinking about it incorrectly for some unknown reason. $\endgroup$ – jbowman Nov 22 '17 at 16:01
1
$\begingroup$

auto.arima() decides on whether or not to use seasonality based on the OCSB test. ets() uses information criteria to test for seasonality or non, as detailed in Hyndman et al., Forecasting with Exponential Smoothing: The State Space Approach. Incidentally, you are right that ETS(M,N,N) is nonseasonal: it has multiplicative errors, no trend and no seasonality.

Given that the two methods use two different approaches to detect seasonality, it's not surprising that the results will differ, unless seasonality is blatantly obvious. And in your case, seasonality is not obvious at all. Compare your data to an obviously seasonal time series:

series

foo <- structure(c(198.5, 206.5, 220.25, 210, 201.5, 235, 248, 249, 
243, 236.25, 236.25, 260.75, 282, 271, 298.8, 351.5, 324.9, 307.8, 
320, 322.6, 334.7, 344.6, 366.3, 395.2, 423.3, 462.4, 435, 440.6, 
412.7, 437.3, 460.9, 453.5, 455.8, 431.4, 424.4, 461.3, 484.7, 
500, 473.7, 435.7, 426.5, 425.2, 440, 320, 349.5, 368.4, 365.7, 
314.7, 329.5, 354.6, 315.9), .Tsp = c(2012.75, 2016.91666666667, 
12), class = "ts")
opar <- par(mai=c(.5,.5,.1,.1),mfrow=c(1,2))
    plot(foo)
    plot(AirPassengers)
par(opar)

One very useful way to detect seasonality is the so-called seasonplot. Here is a seasonplot of your data against the AirPassengers dataset:

seasonplot

library(forecast)
blackBodyRadiationColors <- function(x) {
    # x should be between 0 (black) and 1 (white)
    foo <- colorRamp(c(rgb(0,0,0),rgb(1,0,0),rgb(1,1,0),rgb(1,1,1)))(x)/255
    apply(foo,1,function(bar)rgb(bar[1],bar[2],bar[3]))
}
n.colors <- ceiling(length(foo)/frequency(foo))
colors.blackBody.foo <- blackBodyRadiationColors(seq(0,0.6,length.out=n.colors))
n.colors <- ceiling(length(AirPassengers)/frequency(AirPassengers))
colors.blackBody.AirPassengers <- blackBodyRadiationColors(seq(0,0.6,length.out=n.colors))

opar <- par(mai=c(.5,.5,.5,.1),mfrow=c(1,2))
    seasonplot(foo,col=colors.blackBody.foo,year.labels=TRUE)
    seasonplot(AirPassengers,col=colors.blackBody.AirPassengers,year.labels=TRUE)
par(opar)

It appears obvious to me that there is essentially no seasonal pattern to your data at all. In addition, as jbowman notes, it doesn't really look very stationary, and the large smoothing parameter ets() estimates ($\alpha=0.94$) looks uncomfortably close to a random walk, too. So it may be an idea to first take differences, then potentially look for seasonality.

(Incidentally, I use black body radiation colors because the Rainbow Color Map (is) (Still) Considered Harmful.)

|cite|improve this answer
$\endgroup$
  • $\begingroup$ Many many thanks for the explanation and illustrations. Much clearer now. One last thing, is there any way you would think of to justify the fact that my auto.arima decided to go for seasonality? The forecasts from the auto.arima are very good in my experiment, but if I'm saying that my time series is not showing any seasonal pattern, how can I back up the decision made by auto.arima? $\endgroup$ – Notna Nov 21 '17 at 21:29
  • $\begingroup$ To be honest, I already find it non-trivial to interpret a ARIMA(0,1,0)(1,1,0)[12] model in plain English. (Yes, it's possible.) Remember that auto.arima() uses a statistical test to determine seasonality. Tests do make errors. In this particular case, I would first take differences and then see what happens, and based on the seasonplot, I would be rather skeptical about seasonality here. $\endgroup$ – Stephan Kolassa Nov 21 '17 at 21:35

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.