Removing old implementations from the R package, so will need to grab them from the version history from now on to access the original linear implementations.
Nelder-Mead simplex algorithm seems quite sensitive to starting conditions. Seeding it with values fit from the hansen algorithm (global alpha and sigma), it rapidly improves the log likelihood score (going up from 25 to 30, which is just barely significant according to AIC). With more arbitrary starting conditions it significantly under-performs even Brownian motion models.
Implemented a simulated annealing likelihood maximization routine, which proves much more robust to starting conditions, though taking significantly longer to converge. Also this performance requires appropriately high starting temperature and number of regimes. Hopefully is easy to convert to an MCMC routine. Should add lockable parameters and should use the faster linear solvers for BM and OU1 to make a general MCMC suite for comparative methods.