I am trying to identify changepoints in time series data using this online changepoint detection algorithm. It seems that there are a couple of user defined parameters that I have to give the model:
- A hazard function (I take this to be some way of evaluating how long each period of time steps should be, on average, before we reach a changepoint)
- Some a priori distribution that my data roughly fits, in each partition
My question is about #2. How can I estimate what a good distribution would be to most nearly match my time series for each n steps? I've seen some using using a t distribution and others assuming a normal distribution , and still others using a poisson distribution.
I'm sure my data is not i.i.d, so I can't see how I justify using any of the first two. Is there a more rigorous way to estimate this?