# Does correlation between residuals of ARIMA models for two time series tell anything about how they fluctuate together?

I wonder whether one can judge strength of coupling between fluctuations of two time series by looking at correlation between residuals of ARIMA models for these two series.

Let’s say I have two series, one provides daily air temperature and the second provides water temperatures of a river. Both series are strongly periodic and stationary. I fit, let’s say, ARIMA(2, 0, 0) to both of them and both models are pretty good. Then I check that residuals of both models are also stationary and have no significant autocorrelations. And finally I correlate them using standard Pearson’s $r$ and get a correlation coefficient of about 0.20. Can I say that random fluctuations of one series “explain” about 4% of random fluctuations of another series?

EDIT. The series was periodic (so no stationary). What I mean is that the ARIMA model residuals of both series were stationary.

Cross Validated Asked by sztal on November 14, 2021

Yes, you could look at the correlation between the two residual series, but you should also see if it is really significant. A correlation of 0.2 might well not be, without details (like length of the series) we cannot say.

Better, you should look at the cross-correlation function between the residual series, for instance Cross-correlation significance in R or Do we need to detrend when do Cross-Correlation between two time series?. The correlation you reported probably is the correlation at the same time points, but there could be a lead/lag relationship between the series, and that could show up in the crosscorrelations.

Can I say that random fluctuations of one series "explain" about 4% of random fluctuations of another series?

Maybe not, there could be some common cause behind both, or one could cause the other but from correlations you cannot say which way. Crosscorrelations could help here, if there is either a lead or a lag, maybe the series changing first is driving the other, but also a common cause could act faster on one of the series than on the other ... there are many possibilities, and only correlations do not say enough. Look into multivariate timeseries models like VAR, maybe ...

And, finally:

Both series are strongly periodic and stationary. I fit, let's say, ARIMA(2, 0, 0) to both of them and both models are pretty good

This does NOT exclude stationarity! the periodicity could be dynamic, and AR(2) models does have such quasiperiodic behaviour for some parameter values, and are used precisely to model periodic behaviour. See the answers to Forms of non-stationary process

Answered by kjetil b halvorsen on November 14, 2021

You have stated "both series are strongly periodic and stationary". I guess by "periodic" you meant seasonal.

If a series is seasonal, then it can not be stationary.

Your series and (probably) your residuals are therefore non-stationary and so you can not rely on the correlation coefficient at all.

Answered by Branislav Cuchran on November 14, 2021

## Related Questions

### General procedures for combined feature selection, model tuning, and model selection?

1  Asked on July 31, 2020 by uared1776

### Classification model not working for a large dataset

1  Asked on July 30, 2020 by gabriel-ullmann

### Sigma algebra generated by random variable on a set with generators

0  Asked on July 28, 2020 by gabriel

### What is the seasonal trend lowess model in time series?

0  Asked on July 28, 2020 by christopher-u

### Non seasonal and seasonal parameters of this time-series

0  Asked on July 27, 2020 by statsmonkey

### Extended Cox model and cox.zph

2  Asked on July 25, 2020 by finance