I am reading paper "Bayesian Online Changepoint Detection", which can be found at arxiv:
The results looks inspiring, though it is unclear how to choose (or estimate) the parameters in the priors.
What could be a suitable way to estimate those parameters if one has some historical data?
The problem considered in the paper above is a case of Hidden Markov Models. The authors model the time periods (run lengths) between the changing points of some parameters in the time series.