Lets say i do have a TimeSeries with N > 300 but values like [0, 1, 2, 1, 2, 0, 2, ...] representing the visitors count of a website per day.

Since there are only few visitors and each of them can be considered an individual, is there a way to "prove" or maybe some statement in literature that these values are too low for e.g. prediction with random forest or simply a correlation with other, better performing websites? E.g. if there is a high correlation based on higher visitor counts on mondays, can this correlation actually be considered valid?

Or more specifically: can p < 0.05 as received from scipy.stats.pearsonr actually be considered valid, even if the values of one of the input-arrays are low?

Additionaly, lets say i did some SEO and my visitors count mean improves by 400%, the actual values will still be low and could still be based on random effects, or am i getting this wrong?

Kind regards,


  • $\begingroup$ Please consider that if visits were examined every second rather than every day, for the data suggested the sequence would be thousands of zeros and an occasional one. Finding a correlation for such sparse data would not be very revealing. To do Chi-squared testing, for example, one needs 5 or more counts in each bin, so rebinning the data, e.g., into week by week would make more sense. $\endgroup$ – Carl Sep 20 '18 at 1:35

There's per-se no problem with low values and randomness is a feature of though counts as much as for low counts. However methods for normally distributed data such as tests for Pearson's correlation coefficient are likely very problematic when used for low counts, while for really high counts they may well work. Repeated measures count data models are likely more appropriate.

With counts this low you also may have to worry about this being driven by single individuals and think about the implications of that.


Your Answer

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

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