I’ve recently released a couple of articles on time series forecasting that I want to re-share:
Roughly I am trying to point out alternatives to rushing to ARIMA without trying additional methods.
ARIMA is great at handing the issues of un-observables and serial correlation. My point is: these are not the only issues in time series modeling, and there are additional ways to deal with them. Issues that, in my opinion, ARIMA is not good at include dealing with low signal to noise ratios found in business data and avoiding ill-condition due to short baseline (or short timeframe, ill-conditioned) time series formulation.
As always, go to primary sources (such as Box, Jenkins, Reinsel, Time Series Analysis, 4th Edition Wiley, 2008). Keep in mind: data science is a science, in the sense it requires empirical research.
Categories: Opinion Pragmatic Data Science Tutorials
