1

Has anyone any idea how one could distinguish time series according to certain properties? The only time series properties I know are stationarity/nonstationarity and homoskedasticity/heteroskedasticity. But are there any other possibilities to distinguish time series?

Nick Cox
  • 48,377
  • 8
  • 110
  • 156
DatamineR
  • 1,477
  • 3
  • 18
  • 25
  • Yes, but there really are entire textbooks on this. See http://stats.stackexchange.com/help/dont-ask for guidance on scope. "Your questions should be reasonably scoped. If you can imagine an entire book that answers your question, you’re asking too much." – Nick Cox Jul 11 '13 at 13:04
  • maybe you can recommend me one or two good books? – DatamineR Jul 11 '13 at 13:11
  • Once again, the help explains how you can search for yourself first. http://stats.stackexchange.com/questions/6275/good-introductions-to-time-series-with-r is one of several starting points. – Nick Cox Jul 11 '13 at 13:15
  • There is also this: [books-for-self-studying-time-series-analysis](http://stats.stackexchange.com/questions/20514/). – gung - Reinstate Monica Jul 11 '13 at 13:32
  • I have even read some of the recommended books, but my main interest is how can I group time series according to certain properties, WITHOUT first estimating a model? – DatamineR Jul 11 '13 at 13:37
  • 1
    I am **not** voting to close. Consider this interpretation: "What are the major qualitative properties of time series that one looks at when commencing any analysis/forecast/etc?" Surely our time-series experts would be able to rattle off a half dozen or so characteristics and (I hope) most would agree on them. Offhand, I imagine these would include trend, seasonality, sign of serial correlation, heteroscedasticity, outliers, and any suggestion of changepoints. (It's unclear whether stationarity can be qualitatively apprehended.) A discussion of these could be a useful guide for non-experts. – whuber Jul 11 '13 at 14:45

2 Answers2

1

A very widely used time-series identification scheme is the Box-Jenkins approach. See here. This involves establishing stationarity and decomposing seasonality, then fitting Autoregressive (Integrated) Moving Average (ARIMA) models to the resulting series. The R function ARIMA() in the stats package will do this and chooses the appropriate model and estimates parameters according to AIC.

A point of clarification in your question, however: The concept of stationarity includes the variance (i.e. homoskedasticity vs. heteroskedasticity). A stationary series is, by definition, homoskedastic. See here.

gregory_britten
  • 1,253
  • 9
  • 15
1

I think I have found what I have been looking for in the paper of Rob Hyndman:

"...there are nine classical and advanced statistical features describing a time series’ global characteristics. They are: trend, seasonality, periodicity, serial correlation, skewness, kurtosis, non-linearity, self-similarity, and chaos." (Wang/Smith/Hyndman (2006),Characteristic-Based Clustering for Time Series Data)

DatamineR
  • 1,477
  • 3
  • 18
  • 25