Newbie's questions about timeseries


I have a question which would help me understand better the basics of time-series analysis.

Let’s say that I have a time-series which is simply a sequence of numbers with associated timestamps. I give it to you and I tell you “Study this data and tell me its most important properties” What would you do?


The first thing I’d do is a visual analysis. That helps to understand what data you are dealing with.

Secondly, (if its still not clear), stationarity tests.
If it is stationary (i.e. its just a noise) you dont have much options - just measure mean/std etc.
If it is not, in addition to basic statistics I would check seasonality and trend (check seasonal decomposition), auto-correlation, how well I can approximate it (try different functions), Fourier analysis is an option as well.
Also it’s possible to try all that with log-s or diff-s.

It all depends on the result you expect, in the common case I’d start with the stuff above.


Yep, Alex basically nailed it, I was too slow:)

  • Stationarity first of all, yes - often it’s indeed enough to see the plot to understand this
  • For me since I’ve been working with EM waves I would definitely look at the Fourier Transformation if there is a clear picture from the frequency point of view - but I guess for financial data trends are more important (and trends for the diffs, i.e. accelerating trend, and etc)


After that one could maybe mention AR, MA, ARIMA, SARIMA models as a next step. Those are the “traditional” models from economics to model for when you have only a single time series. With these you can technically already try to predict future prices. The prophet model from facebook that we use for our anomaly detection is basically very similar to these models. Just on steroids ^^