Given a time series (that sometimes unpredictably alternates between random and non-random and is not stationary). What mathematical methods are there to classify the observed set of data into different states?
I have tried classifying the data into states by its empirical distribution for a given period, but find that it is very dependent on parameters such as sample length and bin-size (in histogram). Are there any better methods?
Any suggestions very welcome!
Edit: preferably analytical methods, i.e. no tools such as machine learning
Edit2: The time series has 3 values for each day. (Analogous to measuring temperature but it has no unit). The TS is about 30000 days long (so a 30000x3 matrix). For every day we are given the "temperature" measured at the same time "every day at 12:00", and the highest- / lowest "temperature" value respectively for that same day. My goal is to divide the TS into different states to better visualize the process, maybe find probabilities of going from one state to another (eg. Markov process), and find states that do not communicate etc.
Link to sample of 5000 data points. First column is daily max value, second column is daily minimum value, third is daily measurement at 12:00 and I also added a fourth column (from second csv-file) which is the daily number of observations.