ENH: Multivariate Markov Switching Models

[haginile]

The statespace component already has an excellent module for fitting Markov switching models; however, only univariate models are supported. Is there a way to extend it so as to support multivariate models? Thanks!

[ChadFulton]

We had a GSOC project to allow estimating any Markov switching state space model in #2921, and it is reasonably far along, but it has not been merged (partly because it is written in pure Python and so it can be very slow).

If you mean something like VAR models with Markov switching, those could be estimated using the Hamilton filter and Kim smoother already in Statsmodels (along very similar lines to the markov_autoregression.py file), but we do not have a built-in model to do that yet. This is simply an issue of no one having had the time / energy to write the code, and we would be really happy for any contributions.

[haginile]

Thanks @ChadFulton. I'll take a look and revert back.

[sumit-uk1]

@ChadFulton is there a sample code that you can share? my use case is:-

I have a dataframe of N observable features and I want to fit a HMM Gaussian mixture to that.

Similar to the example in:
https://www.statsmodels.org/devel/examples/notebooks/generated/markov_autoregression.html

I was hoping to get the transition probabilities/ duration of 2 states and state probability predictions

Thanks

[ChadFulton]

@sumit-uk1 I'm not quite sure what model you're looking for:

• If you're looking for a switching regression model with 1 variable on the left-hand-side and N variables on the right-hand-side, then you can use MarkovRegression
• If you're looking for a switching vector autoregression model with N variables on the left-hand-side, then unfortunately we still don't have that built in.

[sumit-uk1]

Hi,   I have a NxM matrix of some returns time series (N<M) Running the code (taken from Markov switching autoregression models — statsmodels) | | | | Markov switching autoregression models — statsmodels | | | ```python    mod_fedfunds = sm.tsa.MarkovRegression(normalized_corr_df, k_regimes=2)res_fedfunds = mod_fedfunds.fit() Gives me the error:ValueError Traceback (most recent call last) <ipython-input-5-200a59bdd86c> in <module>() ----> 1 mod_fedfunds = sm.tsa.MarkovRegression(normalized_corr_df, k_regimes=2) 2 res_fedfunds = mod_fedfunds.fit() ~\Anaconda3\lib\site-packages\statsmodels\tsa\regime_switching\markov_regression.py in __init__(self, endog, k_regimes, trend, exog, order, exog_tvtp, switching_trend, switching_exog, switching_variance, dates, freq, missing) 119 super(MarkovRegression, self).__init__( 120 endog, k_regimes, order=order, exog_tvtp=exog_tvtp, exog=exog, --> 121 dates=dates, freq=freq, missing=missing) 122 123 # Switching options ~\Anaconda3\lib\site-packages\statsmodels\tsa\regime_switching\markov_switching.py in __init__(self, endog, k_regimes, order, exog_tvtp, exog, dates, freq, missing) 513 # Sanity checks 514 if self.endog.ndim > 1 and self.endog.shape[1] > 1: --> 515 raise ValueError('Must have univariate endogenous data.') 516 if self.k_regimes < 2: 517 raise ValueError('Markov switching models must have at least two' ValueError: Must have univariate endogenous data. ``` I was hoping the model will fit a HMM model and give me state transition probabilities and the mean of the 2 states (High vol/ Low vol) RegardsSumit

[ChadFulton]

It sounds like you are looking for a Markov switching VAR model, and we do not support that right now. The only models we have so far are univariate models.