I'm trying to fit a seasonal ARIMA model for a real data time series, 300 samples long, sawtooth looking. I'll write down a resume of my work trying to be clear without many plots.

The ACF of the original data slowly tapering towards 0 and the value PACF(1)=0.8 both suggest a non-seasonal differencing of order 1. As a consequence, both ACF and PACF correlations of the differencied data fall in the 95% confidence interval except for ACF(43)=+0.5 and PACF(~40)=-0.3. This value roughly corresponds to the period of my spikes in the signal.

I therefore applied a seasonal differencing for seasonality=43 multiplied by the non-seas. differencing. The correlograms follow: correlograms

What are the negative correlations at lag=1 telling me, now? Maybe differencing is too much and an AR(1) should be employed?

However, I tried some models with combinations of AR(1), MA(1), SAR(1) and SMA(1) and I got from minimization of AIC and BIC the following best models:

$$(1,0,0)\times(0,1,1)_{43}$$ $$(1,0,1)\times(0,1,0)_{43}$$ $$(1,0,0)\times(0,1,0)_{43}$$

For example, the residuals correlograms for the latter: residuals correlograms for best fit

which look not bad to me even though some significant spikes around the lag of seasonality remain. Can be this due to the imperfect seasonality (data are from extremely fine sampled measurements!)? Anyway, I have other two concerns:

  • the estimated variance of the model is $\sigma=1.7\times 10^{-10}$ with st. error$=4.6\times 10^{-8}$ which gives a t-statistic$=0.002$ really low!

  • the qq-plot of residuals shows a non-agreement with the Gaussian distribution:

qq-plot for best fit

In conclusion, how much can I trust my estimated model given these two problems?


1 Answer 1


Probably not at all ! Your data might disclose changes in parameters over time or deterministic changes in error variance over time suggesting Weighted Least Squares or the need to identify and incorporate latent deterministic structure like pulses, step/level shifts or local time trends OR the need to incorporate seasonal dummies to deal with seasonal structure (often a good choice ! ).

Post your data and let's find out what your data knows.


To review , auto.arima in a brute force list-based procedure that tries a fixed set of models and selects the calculated AIC based upon estimated parameters. The AIC should be calculated from residuals using models that control for intervention administration, otherwise the intervention effects are taken to be Gaussian noise, underestimating the actual model's autoregressive effect and thus miscalculates the model parameters which leads directly to an incorrect error sum of squares and ultimately an incorrect AIC.

The acf of the original is useful in assessing ARIMA structure ( that's the big print ! ) when there are no deterministic factors/features in play like period # ( that's the small print ! ) . Here is the acf of the original data enter image description here .

Your 301 values are indeed non-stationary BUT there is more than 1 remedy for non-stationarity. Your series has deterministic effects for certain periods in the 43 period cycle . See I have correlogram ACF and PACF below for a temperature time series. Can I say it is MA(2) from ACF? What about AR? for a similar but different but related discussion.

I took your 301 values into AUTOBOX a piece of software that I had helped to develop and I asked it to tell me what it could find out by studying the 301 historical values. It hesitated for a very brief amount of time and told me the following.

Here is the Actual /Fit and Forecast graph enter image description here

3: https://i.sstatic.net/lbD0m.png and a less busy Actual/Forecast enter image description here . The USEFUL equation that was found ( Note all models are "rong" but some are useful and some are very wrong ) is here in two parts. enter image description here and enter image description here providing clarity that the arima portion is simply (1,0,0) .

The statistics for the model are presented here enter image description here

The residuals look pristine enter image description here with a very clean acf enter image description here .

A closeup of the forecasts is here ....enter image description here

You said " I therefore applied a seasonal differencing for seasonality=43 " ... and I say that the data says no .

Periods 1-8 AND 36-43 ... are definitely non-players as they have little or no impact on what you are measuring.

It turns out your data had a lot to say about itself ...

With respect to your NON-SIGNIFICANT VARIANCE ... this is probably due to a numerical flaw in the software that you used ...

  • $\begingroup$ Thank you for your reply. Here you can find the raw data: pasted.co/93147588 $\endgroup$
    – LorenzoS
    Commented Jun 27, 2018 at 15:19
  • $\begingroup$ If you like my answer ..please accept it and close the question $\endgroup$
    – IrishStat
    Commented Jun 29, 2018 at 9:32
  • $\begingroup$ thanks for the detailed reply. I'm still trying to fully understand it, but in the meanwhile I accepted your answer. $\endgroup$
    – LorenzoS
    Commented Jun 29, 2018 at 11:17
  • $\begingroup$ If I can help please feel free to contact me via email ...in order to clear up any questions $\endgroup$
    – IrishStat
    Commented Jun 29, 2018 at 11:40
  • $\begingroup$ or open up a new question ...and ask 1 or more so that you can better understand ... ... . If you can't frame the question perhaps dialogue would help. I am interested in what aspect of physics is this ..what were you measuring and how frequently $\endgroup$
    – IrishStat
    Commented Jul 1, 2018 at 13:32

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.