I am writing a thesis which compares two approaches to time series classification: Hidden Markov Models and Dynamic Time Warping combined with 1-NN. I'll apply both algorithms to some real dataset, but first it would be nice to show differences in simulation study.
I have already generated samples from 9 HMM instances created with hmmlearn library (all GaussianHMM, with different parameters) and then compared accuracies of classification using HMM vs DTW. Unsuprisingly HMM wins in this case. Now I'm looking for another model to generate data such that this time DTW yields better results. I've tried ARIMA, but results was awful both for HMM and DTW (but still significantly better for HMM). Which model should I use?