poisson_hmm_constructor.py

    def __init__(self, endog, exog, k_regimes=2, loglike=None, score=None, hessian=None,
                 missing='none', extra_params_names=None, **kwds):
        super(PoissonHMM, self).__init__(endog=endog, exog=exog, loglike=loglike, score=score,
                                         hessian=hessian, missing=missing,
                                         extra_params_names=extra_params_names, kwds=kwds)
        self.y = np.array(self.endog)
        self.k_regimes = k_regimes
        #k_regimes x exog.shape[1] size matrix of regime specific regression coefficients
        self.beta_matrix = np.ones([self.k_regimes, self.exog.shape[1]])
        # k x k matrix of psuedo transition probabilities which can range from -inf to +inf during
        # optimization. Initialized to 1.0/k
        self.q_matrix = np.ones([self.k_regimes,self.k_regimes])*(1.0/self.k_regimes)
        print('self.q_matrix='+str(self.q_matrix))
        #The regime wise matrix of Poisson means. These would be updated during the optimization
        # loop
        self.mu_matrix = []
        # k x k matrix of the real Markov transition probabilities which will be calculated from
        # the q-matrix using a standardization technique. Initialized to 1.0/k
        self.gamma_matrix = np.ones([self.k_regimes, self.k_regimes])*(1.0/self.k_regimes)
        print('self.gamma_matrix='+str(self.gamma_matrix))
        # The Markov state probabilities. Also referred to as pi. but we'll use delta since pi is
        # often used to refer to the mean
        self.delta_matrix = np.ones([self.exog.shape[0],self.k_regimes])*(1.0/self.k_regimes)
        print('self.delta_matrix='+str(self.delta_matrix))
        #The vector of initial values for all the parameters, beta and q, that the optimizer will
        # optimize
        self.start_params = np.repeat(np.ones(self.exog.shape[1]), repeats=self.k_regimes)
        self.start_params = np.append(self.start_params, self.q_matrix.flatten())
        print('self.start_params='+str(self.start_params))
        #A very tiny number (machine specific). Used by the LL function.
        self.EPS = np.MachAr().eps
        #Optimization iteration counter
        self.iter_num=0