I have a time series of monthly temperature data for ~ 100 years. My aim is to turn it into a time series which has neither a trend nor autocorrelation. I want to make sure this is true for the whole time series as well as for the series of each of the months seperately (e.g. looking at a time series of August temperature from 1900-2000). How do I go about that? Especially taking into account that different months might have different trends?

enter image description here enter image description here

  • $\begingroup$ What do you mean by monthly trends? $\endgroup$
    – Tim
    Sep 20 '17 at 9:06
  • $\begingroup$ I mean the trend of one month's temperature over the whole time period (s, second plot) $\endgroup$
    – user45065
    Sep 20 '17 at 9:11
  • $\begingroup$ When you say My aim is to turn it into a time series which has neither a trend nor autocorrelation, what features of the original data would you like to preserve? What would you do with a series once it has been transformed so as to lose some of its defining characteristics? $\endgroup$ Sep 20 '17 at 9:15
  • $\begingroup$ My aim is to compare temperature in different years (same month or mean per year) without having any influence of an underlying trend. $\endgroup$
    – user45065
    Sep 20 '17 at 9:19
  • $\begingroup$ Not quite understood. What variation do you want to concentrate on? Do you want to remove the overall trend to compare variation over the year (i.e. typical January against typical February and the like)? Or do you want to remove the monthly variation to focus on the overall trend (I understand that this is not the case)? Or what else? Please also include @RichardHardy in your responses so that I get notified of them. $\endgroup$ Sep 20 '17 at 10:47

While I don't quite understand what your goal would be, but some general remarks that might be helpful for you:

  1. Removing Trend (DetrendedSignal = Signal - Trend): Judging from the picture you posted, there seems to be a slight positive, linear trend overall. So you might fit a linear model (temp as a function of time) and subtract its estimation from your original data. Of course, you can also try non-linear models..Depending on which software you are using, there surely exist a package/method that does this automatically.

  2. Removing Autocorrelation: One way would be to fit an ARIMA-Model on your data and keep only the residuals (the part of your data that can't be explained via auto-correlation). There are many softwarepackages for this too-

  3. whole time-series vs. separate month: hard to say anything concrete without knowing your application, but wouldn't it be possible to make separate analyses for the whole data and for specific months? If not, you would have to rely on a more complex trend-model, maybe a multiplicative one or seasonal trend decomposition. I found these two tutorials helpful : some general information & seasonal Trend decomposition in R

Hope this was somewhat helpful.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy