0
$\begingroup$

This is somewhat nonstandard:

My input is a series of values (integers or reals), and my goal is to decide if this series is likely to be a series of measurements of some unknown phenomenon (be it an airplane's altitude, a room's temperature, a daily stock price, etc.). If not a series of measurements - this series can be anything.

I have no clue how these values were generated or what do they represent (if anything at all). The only thing I do know is that these values are given in some 'natural' order (temporal, spatial or other).

One test I considered:

  • The average absolute difference between consecutive values is smaller than the average absolute difference between each 2nd element, 3rd element, etc.

    avg(abs((aₙ - aₙ₋₁)) < avg(abs((aₙ - aₙ₋₂)) < avg(abs((aₙ - aₙ₋₃))

I'll be happy to learn if there are known methods to tackle this problem. Alternatively - any idea would be appreciated.

$\endgroup$
  • $\begingroup$ Does your test assume the 'natural' ordering is regular? Many time series can be collected at irregular time points, for example. If you have time stamps, it's easy to check whether or not you have irregularity. $\endgroup$ – Isabella Ghement Oct 27 '18 at 13:07
  • $\begingroup$ I capture these values by monitoring communication channels. In some scenarios, I can assign a timestamp to each value. In other scenarios (e.g. long bursts), I cannot. $\endgroup$ – Lior Kogan Oct 27 '18 at 13:11
  • $\begingroup$ That suggests irregularity then. For things like a daily stock price, I guess you could build a "library" of stocks and then check if what you measured "matches" what you captured. Or maybe you can build a "library of behaviours" for daily stock prices and then compare what you have against that. It seems to me that you would need some external information to identify a "phenomenon" - not just some coarse rule that applies to the series of current measurements (and that may implicitly assume regularity?). $\endgroup$ – Isabella Ghement Oct 27 '18 at 13:20
  • $\begingroup$ Let me put it better: Many scenarios are regular, and a criterion for these is good enough. $\endgroup$ – Lior Kogan Oct 27 '18 at 13:37
2
$\begingroup$

I don't think you will be very successful, though this may depend on your alternative non-time-series data generating processes (which you say you have no clue about).

Yes, you could follow an approach like the one you outline. It is very similar to fitting an model (you could take a look at auto.arima() in the forecast package for R) and declaring a series to be a time series if the resulting model has AR or MA orders >0. Yes, you can do this.

The problem is that correlations between ordered sequences of data can come from different sources than from time orderings. The most common such source may be spatial correlation. Your ordered sequence may be temperatures along a stretch of road, measured simultaneously (so there certainly is no time series aspect to this). Fitting an ARIMA model would reveal strong correlations: temperatures between adjacent measurements are more strongly correlated than between distant measurements. The approach outlined above would believe this is a time series.

$\endgroup$
  • $\begingroup$ This is good enough. Actually, I want to detect if it is likely that these values represent measurements of some phenomenon. I care less if these values are time-series or "space-series". $\endgroup$ – Lior Kogan Oct 27 '18 at 11:25

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.