Source code for statsmodels.tsa.statespace.mlemodel

# -*- coding: utf-8 -*-
"""
State Space Model

Author: Chad Fulton
License: Simplified-BSD
"""
from __future__ import division, absolute_import, print_function
import warnings

import numpy as np
import pandas as pd
from scipy.stats import norm

from statsmodels.compat.python import long

from statsmodels.tools.tools import pinv_extended, Bunch
from statsmodels.tools.sm_exceptions import PrecisionWarning
from statsmodels.tools.numdiff import (_get_epsilon, approx_hess_cs,
                                       approx_fprime_cs, approx_fprime)
from statsmodels.tools.decorators import cache_readonly, resettable_cache
from statsmodels.tools.eval_measures import aic, bic, hqic

import statsmodels.base.wrapper as wrap

import statsmodels.genmod._prediction as pred
from statsmodels.genmod.families.links import identity

import statsmodels.tsa.base.tsa_model as tsbase

from .simulation_smoother import SimulationSmoother
from .kalman_smoother import SmootherResults
from .kalman_filter import INVERT_UNIVARIATE, SOLVE_LU

if bytes != str:
    # PY3
    unicode = str


def _handle_args(names, defaults, *args, **kwargs):
    output_args = []
    # We need to handle positional arguments in two ways, in case this was
    # called by a Scipy optimization routine
    if len(args) > 0:
        # the fit() method will pass a dictionary
        if isinstance(args[0], dict):
            flags = args[0]
        # otherwise, a user may have just used positional arguments...
        else:
            flags = dict(zip(names, args))
        for i in range(len(names)):
            output_args.append(flags.get(names[i], defaults[i]))

        for name, value in flags.items():
            if name in kwargs:
                raise TypeError("loglike() got multiple values for keyword"
                                " argument '%s'" % name)
    else:
        for i in range(len(names)):
            output_args.append(kwargs.pop(names[i], defaults[i]))

    return tuple(output_args) + (kwargs,)


[docs]class MLEModel(tsbase.TimeSeriesModel): r""" State space model for maximum likelihood estimation Parameters ---------- endog : array_like The observed time-series process :math:`y` k_states : int The dimension of the unobserved state process. exog : array_like, optional Array of exogenous regressors, shaped nobs x k. Default is no exogenous regressors. dates : array-like of datetime, optional An array-like object of datetime objects. If a Pandas object is given for endog, it is assumed to have a DateIndex. freq : str, optional The frequency of the time-series. A Pandas offset or 'B', 'D', 'W', 'M', 'A', or 'Q'. This is optional if dates are given. **kwargs Keyword arguments may be used to provide default values for state space matrices or for Kalman filtering options. See `Representation`, and `KalmanFilter` for more details. Attributes ---------- ssm : KalmanFilter Underlying state space representation. Notes ----- This class wraps the state space model with Kalman filtering to add in functionality for maximum likelihood estimation. In particular, it adds the concept of updating the state space representation based on a defined set of parameters, through the `update` method or `updater` attribute (see below for more details on which to use when), and it adds a `fit` method which uses a numerical optimizer to select the parameters that maximize the likelihood of the model. The `start_params` `update` method must be overridden in the child class (and the `transform` and `untransform` methods, if needed). See Also -------- MLEResults statsmodels.tsa.statespace.kalman_filter.KalmanFilter statsmodels.tsa.statespace.representation.Representation """ def __init__(self, endog, k_states, exog=None, dates=None, freq=None, **kwargs): # Initialize the model base super(MLEModel, self).__init__(endog=endog, exog=exog, dates=dates, freq=freq, missing='none') # Store kwargs to recreate model self._init_kwargs = kwargs # Prepared the endog array: C-ordered, shape=(nobs x k_endog) self.endog, self.exog = self.prepare_data() # Dimensions self.nobs = self.endog.shape[0] self.k_states = k_states # Initialize the state-space representation self.initialize_statespace(**kwargs)
[docs] def prepare_data(self): """ Prepare data for use in the state space representation """ endog = np.array(self.data.orig_endog, order='C') exog = self.data.orig_exog if exog is not None: exog = np.array(exog) # Base class may allow 1-dim data, whereas we need 2-dim if endog.ndim == 1: endog.shape = (endog.shape[0], 1) # this will be C-contiguous return endog, exog
[docs] def initialize_statespace(self, **kwargs): """ Initialize the state space representation Parameters ---------- **kwargs Additional keyword arguments to pass to the state space class constructor. """ # (Now self.endog is C-ordered and in long format (nobs x k_endog). To # get F-ordered and in wide format just need to transpose) endog = self.endog.T # Instantiate the state space object self.ssm = SimulationSmoother(endog.shape[0], self.k_states, **kwargs) # Bind the data to the model self.ssm.bind(endog) # Other dimensions, now that `ssm` is available self.k_endog = self.ssm.k_endog
def __setitem__(self, key, value): return self.ssm.__setitem__(key, value) def __getitem__(self, key): return self.ssm.__getitem__(key)
[docs] def set_filter_method(self, filter_method=None, **kwargs): """ Set the filtering method The filtering method controls aspects of which Kalman filtering approach will be used. Parameters ---------- filter_method : integer, optional Bitmask value to set the filter method to. See notes for details. **kwargs Keyword arguments may be used to influence the filter method by setting individual boolean flags. See notes for details. Notes ----- This method is rarely used. See the corresponding function in the `KalmanFilter` class for details. """ self.ssm.set_filter_method(filter_method, **kwargs)
[docs] def set_inversion_method(self, inversion_method=None, **kwargs): """ Set the inversion method The Kalman filter may contain one matrix inversion: that of the forecast error covariance matrix. The inversion method controls how and if that inverse is performed. Parameters ---------- inversion_method : integer, optional Bitmask value to set the inversion method to. See notes for details. **kwargs Keyword arguments may be used to influence the inversion method by setting individual boolean flags. See notes for details. Notes ----- This method is rarely used. See the corresponding function in the `KalmanFilter` class for details. """ self.ssm.set_inversion_method(inversion_method, **kwargs)
[docs] def set_stability_method(self, stability_method=None, **kwargs): """ Set the numerical stability method The Kalman filter is a recursive algorithm that may in some cases suffer issues with numerical stability. The stability method controls what, if any, measures are taken to promote stability. Parameters ---------- stability_method : integer, optional Bitmask value to set the stability method to. See notes for details. **kwargs Keyword arguments may be used to influence the stability method by setting individual boolean flags. See notes for details. Notes ----- This method is rarely used. See the corresponding function in the `KalmanFilter` class for details. """ self.ssm.set_stability_method(stability_method, **kwargs)
[docs] def set_conserve_memory(self, conserve_memory=None, **kwargs): """ Set the memory conservation method By default, the Kalman filter computes a number of intermediate matrices at each iteration. The memory conservation options control which of those matrices are stored. Parameters ---------- conserve_memory : integer, optional Bitmask value to set the memory conservation method to. See notes for details. **kwargs Keyword arguments may be used to influence the memory conservation method by setting individual boolean flags. Notes ----- This method is rarely used. See the corresponding function in the `KalmanFilter` class for details. """ self.ssm.set_conserve_memory(conserve_memory, **kwargs)
[docs] def set_smoother_output(self, smoother_output=None, **kwargs): """ Set the smoother output The smoother can produce several types of results. The smoother output variable controls which are calculated and returned. Parameters ---------- smoother_output : integer, optional Bitmask value to set the smoother output to. See notes for details. **kwargs Keyword arguments may be used to influence the smoother output by setting individual boolean flags. Notes ----- This method is rarely used. See the corresponding function in the `KalmanSmoother` class for details. """ self.ssm.set_smoother_output(smoother_output, **kwargs)
[docs] def initialize_known(self, initial_state, initial_state_cov): self.ssm.initialize_known(initial_state, initial_state_cov)
[docs] def initialize_approximate_diffuse(self, variance=None): self.ssm.initialize_approximate_diffuse(variance)
[docs] def initialize_stationary(self): self.ssm.initialize_stationary()
@property def initialization(self): return self.ssm.initialization @initialization.setter def initialization(self, value): self.ssm.initialization = value @property def initial_variance(self): return self.ssm.initial_variance @initial_variance.setter def initial_variance(self, value): self.ssm.initial_variance = value @property def loglikelihood_burn(self): return self.ssm.loglikelihood_burn @loglikelihood_burn.setter def loglikelihood_burn(self, value): self.ssm.loglikelihood_burn = value @property def tolerance(self): return self.ssm.tolerance @tolerance.setter def tolerance(self, value): self.ssm.tolerance = value
[docs] def fit(self, start_params=None, transformed=True, cov_type='opg', cov_kwds=None, method='lbfgs', maxiter=50, full_output=1, disp=5, callback=None, return_params=False, optim_score=None, optim_complex_step=None, optim_hessian=None, flags=None, **kwargs): """ Fits the model by maximum likelihood via Kalman filter. Parameters ---------- start_params : array_like, optional Initial guess of the solution for the loglikelihood maximization. If None, the default is given by Model.start_params. transformed : boolean, optional Whether or not `start_params` is already transformed. Default is True. cov_type : str, optional The `cov_type` keyword governs the method for calculating the covariance matrix of parameter estimates. Can be one of: - 'opg' for the outer product of gradient estimator - 'oim' for the observed information matrix estimator, calculated using the method of Harvey (1989) - 'approx' for the observed information matrix estimator, calculated using a numerical approximation of the Hessian matrix. - 'robust' for an approximate (quasi-maximum likelihood) covariance matrix that may be valid even in the presense of some misspecifications. Intermediate calculations use the 'oim' method. - 'robust_approx' is the same as 'robust' except that the intermediate calculations use the 'approx' method. - 'none' for no covariance matrix calculation. cov_kwds : dict or None, optional A dictionary of arguments affecting covariance matrix computation. **opg, oim, approx, robust, robust_approx** - 'approx_complex_step' : boolean, optional - If True, numerical approximations are computed using complex-step methods. If False, numerical approximations are computed using finite difference methods. Default is True. - 'approx_centered' : boolean, optional - If True, numerical approximations computed using finite difference methods use a centered approximation. Default is False. method : str, optional The `method` determines which solver from `scipy.optimize` is used, and it can be chosen from among the following strings: - 'newton' for Newton-Raphson, 'nm' for Nelder-Mead - 'bfgs' for Broyden-Fletcher-Goldfarb-Shanno (BFGS) - 'lbfgs' for limited-memory BFGS with optional box constraints - 'powell' for modified Powell's method - 'cg' for conjugate gradient - 'ncg' for Newton-conjugate gradient - 'basinhopping' for global basin-hopping solver The explicit arguments in `fit` are passed to the solver, with the exception of the basin-hopping solver. Each solver has several optional arguments that are not the same across solvers. See the notes section below (or scipy.optimize) for the available arguments and for the list of explicit arguments that the basin-hopping solver supports. maxiter : int, optional The maximum number of iterations to perform. full_output : boolean, optional Set to True to have all available output in the Results object's mle_retvals attribute. The output is dependent on the solver. See LikelihoodModelResults notes section for more information. disp : boolean, optional Set to True to print convergence messages. callback : callable callback(xk), optional Called after each iteration, as callback(xk), where xk is the current parameter vector. return_params : boolean, optional Whether or not to return only the array of maximizing parameters. Default is False. optim_score : {'harvey', 'approx'} or None, optional The method by which the score vector is calculated. 'harvey' uses the method from Harvey (1989), 'approx' uses either finite difference or complex step differentiation depending upon the value of `optim_complex_step`, and None uses the built-in gradient approximation of the optimizer. Default is None. This keyword is only relevant if the optimization method uses the score. optim_complex_step : bool, optional Whether or not to use complex step differentiation when approximating the score; if False, finite difference approximation is used. Default is True. This keyword is only relevant if `optim_score` is set to 'harvey' or 'approx'. optim_hessian : {'opg','oim','approx'}, optional The method by which the Hessian is numerically approximated. 'opg' uses outer product of gradients, 'oim' uses the information matrix formula from Harvey (1989), and 'approx' uses numerical approximation. This keyword is only relevant if the optimization method uses the Hessian matrix. **kwargs Additional keyword arguments to pass to the optimizer. Returns ------- MLEResults See also -------- statsmodels.base.model.LikelihoodModel.fit MLEResults """ if start_params is None: start_params = self.start_params transformed = True # Update the score method if optim_score is None and method == 'lbfgs': kwargs.setdefault('approx_grad', True) kwargs.setdefault('epsilon', 1e-5) elif optim_score is None: optim_score = 'approx' # Check for complex step differentiation if optim_complex_step is None: optim_complex_step = not self.ssm._complex_endog elif optim_complex_step and self.ssm._complex_endog: raise ValueError('Cannot use complex step derivatives when data' ' or parameters are complex.') # Unconstrain the starting parameters if transformed: start_params = self.untransform_params(np.array(start_params)) # Maximum likelihood estimation if flags is None: flags = {} flags.update({ 'transformed': False, 'score_method': optim_score, 'approx_complex_step': optim_complex_step }) if optim_hessian is not None: flags['hessian_method'] = optim_hessian fargs = (flags,) mlefit = super(MLEModel, self).fit(start_params, method=method, fargs=fargs, maxiter=maxiter, full_output=full_output, disp=disp, callback=callback, skip_hessian=True, **kwargs) # Just return the fitted parameters if requested if return_params: return self.transform_params(mlefit.params) # Otherwise construct the results class if desired else: res = self.smooth(mlefit.params, transformed=False, cov_type=cov_type, cov_kwds=cov_kwds) res.mlefit = mlefit res.mle_retvals = mlefit.mle_retvals res.mle_settings = mlefit.mle_settings return res
@property def _res_classes(self): return {'fit': (MLEResults, MLEResultsWrapper)} def _wrap_results(self, params, result, return_raw, cov_type=None, cov_kwds=None, results_class=None, wrapper_class=None): if not return_raw: # Wrap in a results object result_kwargs = {} if cov_type is not None: result_kwargs['cov_type'] = cov_type if cov_kwds is not None: result_kwargs['cov_kwds'] = cov_kwds if results_class is None: results_class = self._res_classes['fit'][0] if wrapper_class is None: wrapper_class = self._res_classes['fit'][1] res = results_class(self, params, result, **result_kwargs) result = wrapper_class(res) return result
[docs] def filter(self, params, transformed=True, complex_step=False, cov_type=None, cov_kwds=None, return_ssm=False, results_class=None, results_wrapper_class=None, **kwargs): """ Kalman filtering Parameters ---------- params : array_like Array of parameters at which to evaluate the loglikelihood function. transformed : boolean, optional Whether or not `params` is already transformed. Default is True. return_ssm : boolean,optional Whether or not to return only the state space output or a full results object. Default is to return a full results object. cov_type : str, optional See `MLEResults.fit` for a description of covariance matrix types for results object. cov_kwds : dict or None, optional See `MLEResults.get_robustcov_results` for a description required keywords for alternative covariance estimators **kwargs Additional keyword arguments to pass to the Kalman filter. See `KalmanFilter.filter` for more details. """ params = np.array(params, ndmin=1) if not transformed: params = self.transform_params(params) self.update(params, transformed=True, complex_step=complex_step) # Save the parameter names self.data.param_names = self.param_names if complex_step: kwargs['inversion_method'] = INVERT_UNIVARIATE | SOLVE_LU # Get the state space output result = self.ssm.filter(complex_step=complex_step, **kwargs) # Wrap in a results object return self._wrap_results(params, result, return_ssm, cov_type, cov_kwds, results_class, results_wrapper_class)
[docs] def smooth(self, params, transformed=True, complex_step=False, cov_type=None, cov_kwds=None, return_ssm=False, results_class=None, results_wrapper_class=None, **kwargs): """ Kalman smoothing Parameters ---------- params : array_like Array of parameters at which to evaluate the loglikelihood function. transformed : boolean, optional Whether or not `params` is already transformed. Default is True. return_ssm : boolean,optional Whether or not to return only the state space output or a full results object. Default is to return a full results object. cov_type : str, optional See `MLEResults.fit` for a description of covariance matrix types for results object. cov_kwds : dict or None, optional See `MLEResults.get_robustcov_results` for a description required keywords for alternative covariance estimators **kwargs Additional keyword arguments to pass to the Kalman filter. See `KalmanFilter.filter` for more details. """ params = np.array(params, ndmin=1) if not transformed: params = self.transform_params(params) self.update(params, transformed=True, complex_step=complex_step) # Save the parameter names self.data.param_names = self.param_names if complex_step: kwargs['inversion_method'] = INVERT_UNIVARIATE | SOLVE_LU # Get the state space output result = self.ssm.smooth(complex_step=complex_step, **kwargs) # Wrap in a results object return self._wrap_results(params, result, return_ssm, cov_type, cov_kwds, results_class, results_wrapper_class)
_loglike_param_names = ['transformed', 'complex_step'] _loglike_param_defaults = [True, False]
[docs] def loglike(self, params, *args, **kwargs): """ Loglikelihood evaluation Parameters ---------- params : array_like Array of parameters at which to evaluate the loglikelihood function. transformed : boolean, optional Whether or not `params` is already transformed. Default is True. kwargs Additional keyword arguments to pass to the Kalman filter. See `KalmanFilter.filter` for more details. Notes ----- [1]_ recommend maximizing the average likelihood to avoid scale issues; this is done automatically by the base Model fit method. References ---------- .. [1] Koopman, Siem Jan, Neil Shephard, and Jurgen A. Doornik. 1999. Statistical Algorithms for Models in State Space Using SsfPack 2.2. Econometrics Journal 2 (1): 107-60. doi:10.1111/1368-423X.00023. See Also -------- update : modifies the internal state of the state space model to reflect new params """ transformed, complex_step, kwargs = _handle_args( MLEModel._loglike_param_names, MLEModel._loglike_param_defaults, *args, **kwargs) if not transformed: params = self.transform_params(params) self.update(params, transformed=True, complex_step=complex_step) if complex_step: kwargs['inversion_method'] = INVERT_UNIVARIATE | SOLVE_LU loglike = self.ssm.loglike(complex_step=complex_step, **kwargs) # Koopman, Shephard, and Doornik recommend maximizing the average # likelihood to avoid scale issues, but the averaging is done # automatically in the base model `fit` method return loglike
[docs] def loglikeobs(self, params, transformed=True, complex_step=False, **kwargs): """ Loglikelihood evaluation Parameters ---------- params : array_like Array of parameters at which to evaluate the loglikelihood function. transformed : boolean, optional Whether or not `params` is already transformed. Default is True. **kwargs Additional keyword arguments to pass to the Kalman filter. See `KalmanFilter.filter` for more details. Notes ----- [1]_ recommend maximizing the average likelihood to avoid scale issues; this is done automatically by the base Model fit method. References ---------- .. [1] Koopman, Siem Jan, Neil Shephard, and Jurgen A. Doornik. 1999. Statistical Algorithms for Models in State Space Using SsfPack 2.2. Econometrics Journal 2 (1): 107-60. doi:10.1111/1368-423X.00023. See Also -------- update : modifies the internal state of the Model to reflect new params """ if not transformed: params = self.transform_params(params) # If we're using complex-step differentiation, then we can't use # Cholesky factorization if complex_step: kwargs['inversion_method'] = INVERT_UNIVARIATE | SOLVE_LU self.update(params, transformed=True, complex_step=complex_step) return self.ssm.loglikeobs(complex_step=complex_step, **kwargs)
[docs] def simulation_smoother(self, simulation_output=None, **kwargs): r""" Retrieve a simulation smoother for the state space model. Parameters ---------- simulation_output : int, optional Determines which simulation smoother output is calculated. Default is all (including state and disturbances). **kwargs Additional keyword arguments, used to set the simulation output. See `set_simulation_output` for more details. Returns ------- SimulationSmoothResults """ return self.ssm.simulation_smoother( simulation_output=simulation_output, **kwargs)
def _forecasts_error_partial_derivatives(self, params, transformed=True, approx_complex_step=None, approx_centered=False, res=None, **kwargs): params = np.array(params, ndmin=1) # We can't use complex-step differentiation with non-transformed # parameters if approx_complex_step is None: approx_complex_step = transformed if not transformed and approx_complex_step: raise ValueError("Cannot use complex-step approximations to" " calculate the observed_information_matrix" " with untransformed parameters.") # If we're using complex-step differentiation, then we can't use # Cholesky factorization if approx_complex_step: kwargs['inversion_method'] = INVERT_UNIVARIATE | SOLVE_LU # Get values at the params themselves if res is None: self.update(params, transformed=transformed, complex_step=approx_complex_step) res = self.ssm.filter(complex_step=approx_complex_step, **kwargs) # Setup n = len(params) # Compute partial derivatives w.r.t. forecast error and forecast # error covariance partials_forecasts_error = ( np.zeros((self.k_endog, self.nobs, n)) ) partials_forecasts_error_cov = ( np.zeros((self.k_endog, self.k_endog, self.nobs, n)) ) if approx_complex_step: epsilon = _get_epsilon(params, 2, None, n) increments = np.identity(n) * 1j * epsilon for i, ih in enumerate(increments): self.update(params + ih, transformed=transformed, complex_step=True) _res = self.ssm.filter(complex_step=True, **kwargs) partials_forecasts_error[:, :, i] = ( _res.forecasts_error.imag / epsilon[i] ) partials_forecasts_error_cov[:, :, :, i] = ( _res.forecasts_error_cov.imag / epsilon[i] ) elif not approx_centered: epsilon = _get_epsilon(params, 2, None, n) ei = np.zeros((n,), float) for i in range(n): ei[i] = epsilon[i] self.update(params + ei, transformed=transformed, complex_step=False) _res = self.ssm.filter(complex_step=False, **kwargs) partials_forecasts_error[:, :, i] = ( _res.forecasts_error - res.forecasts_error) / epsilon[i] partials_forecasts_error_cov[:, :, :, i] = ( _res.forecasts_error_cov - res.forecasts_error_cov) / epsilon[i] ei[i] = 0.0 else: epsilon = _get_epsilon(params, 3, None, n) / 2. ei = np.zeros((n,), float) for i in range(n): ei[i] = epsilon[i] self.update(params + ei, transformed=transformed, complex_step=False) _res1 = self.ssm.filter(complex_step=False, **kwargs) self.update(params - ei, transformed=transformed, complex_step=False) _res2 = self.ssm.filter(complex_step=False, **kwargs) partials_forecasts_error[:, :, i] = ( (_res1.forecasts_error - _res2.forecasts_error) / (2 * epsilon[i])) partials_forecasts_error_cov[:, :, :, i] = ( (_res1.forecasts_error_cov - _res2.forecasts_error_cov) / (2 * epsilon[i])) ei[i] = 0.0 return partials_forecasts_error, partials_forecasts_error_cov
[docs] def observed_information_matrix(self, params, transformed=True, approx_complex_step=None, approx_centered=False, **kwargs): """ Observed information matrix Parameters ---------- params : array_like, optional Array of parameters at which to evaluate the loglikelihood function. **kwargs Additional keyword arguments to pass to the Kalman filter. See `KalmanFilter.filter` for more details. Notes ----- This method is from Harvey (1989), which shows that the information matrix only depends on terms from the gradient. This implementation is partially analytic and partially numeric approximation, therefore, because it uses the analytic formula for the information matrix, with numerically computed elements of the gradient. References ---------- Harvey, Andrew C. 1990. Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge University Press. """ params = np.array(params, ndmin=1) # Setup n = len(params) # We can't use complex-step differentiation with non-transformed # parameters if approx_complex_step is None: approx_complex_step = transformed if not transformed and approx_complex_step: raise ValueError("Cannot use complex-step approximations to" " calculate the observed_information_matrix" " with untransformed parameters.") # Get values at the params themselves self.update(params, transformed=transformed, complex_step=approx_complex_step) # If we're using complex-step differentiation, then we can't use # Cholesky factorization if approx_complex_step: kwargs['inversion_method'] = INVERT_UNIVARIATE | SOLVE_LU res = self.ssm.filter(complex_step=approx_complex_step, **kwargs) dtype = self.ssm.dtype # Save this for inversion later inv_forecasts_error_cov = res.forecasts_error_cov.copy() partials_forecasts_error, partials_forecasts_error_cov = ( self._forecasts_error_partial_derivatives( params, transformed=transformed, approx_complex_step=approx_complex_step, approx_centered=approx_centered, res=res, **kwargs)) # Compute the information matrix tmp = np.zeros((self.k_endog, self.k_endog, self.nobs, n), dtype=dtype) information_matrix = np.zeros((n, n), dtype=dtype) for t in range(self.ssm.loglikelihood_burn, self.nobs): inv_forecasts_error_cov[:, :, t] = ( np.linalg.inv(res.forecasts_error_cov[:, :, t]) ) for i in range(n): tmp[:, :, t, i] = np.dot( inv_forecasts_error_cov[:, :, t], partials_forecasts_error_cov[:, :, t, i] ) for i in range(n): for j in range(n): information_matrix[i, j] += ( 0.5 * np.trace(np.dot(tmp[:, :, t, i], tmp[:, :, t, j])) ) information_matrix[i, j] += np.inner( partials_forecasts_error[:, t, i], np.dot(inv_forecasts_error_cov[:, :, t], partials_forecasts_error[:, t, j]) ) return information_matrix / (self.nobs - self.ssm.loglikelihood_burn)
[docs] def opg_information_matrix(self, params, transformed=True, approx_complex_step=None, **kwargs): """ Outer product of gradients information matrix Parameters ---------- params : array_like, optional Array of parameters at which to evaluate the loglikelihood function. **kwargs Additional arguments to the `loglikeobs` method. References ---------- Berndt, Ernst R., Bronwyn Hall, Robert Hall, and Jerry Hausman. 1974. Estimation and Inference in Nonlinear Structural Models. NBER Chapters. National Bureau of Economic Research, Inc. """ # We can't use complex-step differentiation with non-transformed # parameters if approx_complex_step is None: approx_complex_step = transformed if not transformed and approx_complex_step: raise ValueError("Cannot use complex-step approximations to" " calculate the observed_information_matrix" " with untransformed parameters.") score_obs = self.score_obs(params, transformed=transformed, approx_complex_step=approx_complex_step, **kwargs).transpose() return ( np.inner(score_obs, score_obs) / (self.nobs - self.ssm.loglikelihood_burn) )
def _score_complex_step(self, params, **kwargs): # the default epsilon can be too small # inversion_method = INVERT_UNIVARIATE | SOLVE_LU epsilon = _get_epsilon(params, 2., None, len(params)) kwargs['transformed'] = True kwargs['complex_step'] = True return approx_fprime_cs(params, self.loglike, epsilon=epsilon, kwargs=kwargs) def _score_finite_difference(self, params, approx_centered=False, **kwargs): kwargs['transformed'] = True return approx_fprime(params, self.loglike, kwargs=kwargs, centered=approx_centered) def _score_harvey(self, params, approx_complex_step=True, **kwargs): score_obs = self._score_obs_harvey( params, approx_complex_step=approx_complex_step, **kwargs) return np.sum(score_obs, axis=0) def _score_obs_harvey(self, params, approx_complex_step=True, approx_centered=False, **kwargs): """ Score Parameters ---------- params : array_like, optional Array of parameters at which to evaluate the loglikelihood function. **kwargs Additional keyword arguments to pass to the Kalman filter. See `KalmanFilter.filter` for more details. Notes ----- This method is from Harvey (1989), section 3.4.5 References ---------- Harvey, Andrew C. 1990. Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge University Press. """ params = np.array(params, ndmin=1) n = len(params) # Get values at the params themselves self.update(params, transformed=True, complex_step=approx_complex_step) if approx_complex_step: kwargs['inversion_method'] = INVERT_UNIVARIATE | SOLVE_LU res = self.ssm.filter(complex_step=approx_complex_step, **kwargs) # Get forecasts error partials partials_forecasts_error, partials_forecasts_error_cov = ( self._forecasts_error_partial_derivatives( params, transformed=True, approx_complex_step=approx_complex_step, approx_centered=approx_centered, res=res, **kwargs)) # Compute partial derivatives w.r.t. likelihood function partials = np.zeros((self.nobs, n)) k_endog = self.k_endog for t in range(self.nobs): inv_forecasts_error_cov = np.linalg.inv( res.forecasts_error_cov[:, :, t]) for i in range(n): partials[t, i] += np.trace(np.dot( np.dot(inv_forecasts_error_cov, partials_forecasts_error_cov[:, :, t, i]), (np.eye(k_endog) - np.dot(inv_forecasts_error_cov, np.outer(res.forecasts_error[:, t], res.forecasts_error[:, t]))))) # 2 * dv / di * F^{-1} v_t # where x = F^{-1} v_t or F x = v partials[t, i] += 2 * np.dot( partials_forecasts_error[:, t, i], np.dot(inv_forecasts_error_cov, res.forecasts_error[:, t])) return -partials / 2. _score_param_names = ['transformed', 'score_method', 'approx_complex_step', 'approx_centered'] _score_param_defaults = [True, 'approx', None, False]
[docs] def score(self, params, *args, **kwargs): """ Compute the score function at params. Parameters ---------- params : array_like Array of parameters at which to evaluate the score. args Additional positional arguments to the `loglike` method. kwargs Additional keyword arguments to the `loglike` method. Returns ---------- score : array Score, evaluated at `params`. Notes ----- This is a numerical approximation, calculated using first-order complex step differentiation on the `loglike` method. Both \*args and \*\*kwargs are necessary because the optimizer from `fit` must call this function and only supports passing arguments via \*args (for example `scipy.optimize.fmin_l_bfgs`). """ params = np.array(params, ndmin=1) transformed, method, approx_complex_step, approx_centered, kwargs = ( _handle_args(MLEModel._score_param_names, MLEModel._score_param_defaults, *args, **kwargs)) # For fit() calls, the method is called 'score_method' (to distinguish # it from the method used for fit) but generally in kwargs the method # will just be called 'method' if 'method' in kwargs: method = kwargs.pop('method') if approx_complex_step is None: approx_complex_step = not self.ssm._complex_endog if approx_complex_step and self.ssm._complex_endog: raise ValueError('Cannot use complex step derivatives when data' ' or parameters are complex.') if not transformed: transform_score = self.transform_jacobian(params) params = self.transform_params(params) if method == 'harvey': score = self._score_harvey( params, approx_complex_step=approx_complex_step, **kwargs) elif method == 'approx' and approx_complex_step: score = self._score_complex_step(params, **kwargs) elif method == 'approx': score = self._score_finite_difference( params, approx_centered=approx_centered, **kwargs) else: raise NotImplementedError('Invalid score method.') if not transformed: score = np.dot(transform_score, score) return score
[docs] def score_obs(self, params, method='approx', transformed=True, approx_complex_step=None, approx_centered=False, **kwargs): """ Compute the score per observation, evaluated at params Parameters ---------- params : array_like Array of parameters at which to evaluate the score. kwargs Additional arguments to the `loglike` method. Returns ---------- score : array Score per observation, evaluated at `params`. Notes ----- This is a numerical approximation, calculated using first-order complex step differentiation on the `loglikeobs` method. """ params = np.array(params, ndmin=1) if not transformed and approx_complex_step: raise ValueError("Cannot use complex-step approximations to" " calculate the score at each observation" " with untransformed parameters.") if approx_complex_step is None: approx_complex_step = not self.ssm._complex_endog if approx_complex_step and self.ssm._complex_endog: raise ValueError('Cannot use complex step derivatives when data' ' or parameters are complex.') if method == 'harvey': score = self._score_obs_harvey( params, transformed=transformed, approx_complex_step=approx_complex_step, **kwargs) elif method == 'approx' and approx_complex_step: # the default epsilon can be too small epsilon = _get_epsilon(params, 2., None, len(params)) kwargs['complex_step'] = True kwargs['transformed'] = True score = approx_fprime_cs(params, self.loglikeobs, epsilon=epsilon, kwargs=kwargs) elif method == 'approx': kwargs['transformed'] = transformed score = approx_fprime(params, self.loglikeobs, kwargs=kwargs, centered=approx_centered) else: raise NotImplementedError('Invalid scoreobs method.') return score
_hessian_param_names = ['transformed', 'hessian_method', 'approx_complex_step', 'approx_centered'] _hessian_param_defaults = [True, 'approx', None, False]
[docs] def hessian(self, params, *args, **kwargs): r""" Hessian matrix of the likelihood function, evaluated at the given parameters Parameters ---------- params : array_like Array of parameters at which to evaluate the hessian. args Additional positional arguments to the `loglike` method. kwargs Additional keyword arguments to the `loglike` method. Returns ------- hessian : array Hessian matrix evaluated at `params` Notes ----- This is a numerical approximation. Both \*args and \*\*kwargs are necessary because the optimizer from `fit` must call this function and only supports passing arguments via \*args (for example `scipy.optimize.fmin_l_bfgs`). """ transformed, method, approx_complex_step, approx_centered, kwargs = ( _handle_args(MLEModel._hessian_param_names, MLEModel._hessian_param_defaults, *args, **kwargs)) # For fit() calls, the method is called 'hessian_method' (to # distinguish it from the method used for fit) but generally in kwargs # the method will just be called 'method' if 'method' in kwargs: method = kwargs.pop('method') if not transformed and approx_complex_step: raise ValueError("Cannot use complex-step approximations to" " calculate the hessian with untransformed" " parameters.") if approx_complex_step is None: approx_complex_step = not self.ssm._complex_endog if approx_complex_step and self.ssm._complex_endog: raise ValueError('Cannot use complex step derivatives when data' ' or parameters are complex.') if method == 'oim': hessian = self._hessian_oim( params, transformed=transformed, approx_complex_step=approx_complex_step, approx_centered=approx_centered, **kwargs) elif method == 'opg': hessian = self._hessian_opg( params, transformed=transformed, approx_complex_step=approx_complex_step, approx_centered=approx_centered, **kwargs) elif method == 'approx' and approx_complex_step: hessian = self._hessian_complex_step( params, transformed=transformed, **kwargs) elif method == 'approx': hessian = self._hessian_finite_difference( params, transformed=transformed, approx_centered=approx_centered, **kwargs) else: raise NotImplementedError('Invalid Hessian calculation method.') return hessian
def _hessian_oim(self, params, **kwargs): """ Hessian matrix computed using the Harvey (1989) information matrix """ return -self.observed_information_matrix(params, **kwargs) def _hessian_opg(self, params, **kwargs): """ Hessian matrix computed using the outer product of gradients information matrix """ return -self.opg_information_matrix(params, **kwargs) def _hessian_finite_difference(self, params, approx_centered=False, **kwargs): params = np.array(params, ndmin=1) warnings.warn('Calculation of the Hessian using finite differences' ' is usually subject to substantial approximation' ' errors.', PrecisionWarning) if not approx_centered: epsilon = _get_epsilon(params, 3, None, len(params)) else: epsilon = _get_epsilon(params, 4, None, len(params)) / 2 hessian = approx_fprime(params, self._score_finite_difference, epsilon=epsilon, kwargs=kwargs, centered=approx_centered) return hessian / (self.nobs - self.ssm.loglikelihood_burn) def _hessian_complex_step(self, params, **kwargs): """ Hessian matrix computed by second-order complex-step differentiation on the `loglike` function. """ # the default epsilon can be too small epsilon = _get_epsilon(params, 3., None, len(params)) kwargs['transformed'] = True kwargs['complex_step'] = True hessian = approx_hess_cs( params, self.loglike, epsilon=epsilon, kwargs=kwargs) return hessian / (self.nobs - self.ssm.loglikelihood_burn) @property def start_params(self): """ (array) Starting parameters for maximum likelihood estimation. """ if hasattr(self, '_start_params'): return self._start_params else: raise NotImplementedError @property def param_names(self): """ (list of str) List of human readable parameter names (for parameters actually included in the model). """ if hasattr(self, '_param_names'): return self._param_names else: try: names = ['param.%d' % i for i in range(len(self.start_params))] except NotImplementedError: names = [] return names
[docs] def transform_jacobian(self, unconstrained, approx_centered=False): """ Jacobian matrix for the parameter transformation function Parameters ---------- unconstrained : array_like Array of unconstrained parameters used by the optimizer. Returns ------- jacobian : array Jacobian matrix of the transformation, evaluated at `unconstrained` Notes ----- This is a numerical approximation using finite differences. Note that in general complex step methods cannot be used because it is not guaranteed that the `transform_params` method is a real function (e.g. if Cholesky decomposition is used). See Also -------- transform_params """ return approx_fprime(unconstrained, self.transform_params, centered=approx_centered)
[docs] def transform_params(self, unconstrained): """ Transform unconstrained parameters used by the optimizer to constrained parameters used in likelihood evaluation Parameters ---------- unconstrained : array_like Array of unconstrained parameters used by the optimizer, to be transformed. Returns ------- constrained : array_like Array of constrained parameters which may be used in likelihood evalation. Notes ----- This is a noop in the base class, subclasses should override where appropriate. """ return np.array(unconstrained, ndmin=1)
[docs] def untransform_params(self, constrained): """ Transform constrained parameters used in likelihood evaluation to unconstrained parameters used by the optimizer Parameters ---------- constrained : array_like Array of constrained parameters used in likelihood evalution, to be transformed. Returns ------- unconstrained : array_like Array of unconstrained parameters used by the optimizer. Notes ----- This is a noop in the base class, subclasses should override where appropriate. """ return np.array(constrained, ndmin=1)
[docs] def update(self, params, transformed=True, complex_step=False): """ Update the parameters of the model Parameters ---------- params : array_like Array of new parameters. transformed : boolean, optional Whether or not `params` is already transformed. If set to False, `transform_params` is called. Default is True. Returns ------- params : array_like Array of parameters. Notes ----- Since Model is a base class, this method should be overridden by subclasses to perform actual updating steps. """ params = np.array(params, ndmin=1) if not transformed: params = self.transform_params(params) return params
[docs] def simulate(self, params, nsimulations, measurement_shocks=None, state_shocks=None, initial_state=None): r""" Simulate a new time series following the state space model Parameters ---------- params : array_like Array of model parameters. nsimulations : int The number of observations to simulate. If the model is time-invariant this can be any number. If the model is time-varying, then this number must be less than or equal to the number measurement_shocks : array_like, optional If specified, these are the shocks to the measurement equation, :math:`\varepsilon_t`. If unspecified, these are automatically generated using a pseudo-random number generator. If specified, must be shaped `nsimulations` x `k_endog`, where `k_endog` is the same as in the state space model. state_shocks : array_like, optional If specified, these are the shocks to the state equation, :math:`\eta_t`. If unspecified, these are automatically generated using a pseudo-random number generator. If specified, must be shaped `nsimulations` x `k_posdef` where `k_posdef` is the same as in the state space model. initial_state : array_like, optional If specified, this is the state vector at time zero, which should be shaped (`k_states` x 1), where `k_states` is the same as in the state space model. If unspecified, but the model has been initialized, then that initialization is used. If unspecified and the model has not been initialized, then a vector of zeros is used. Note that this is not included in the returned `simulated_states` array. Returns ------- simulated_obs : array An (nsimulations x k_endog) array of simulated observations. """ self.update(params) simulated_obs, simulated_states = self.ssm.simulate( nsimulations, measurement_shocks, state_shocks, initial_state) # Simulated obs is (nobs x k_endog); don't want to squeeze in # case of nsimulations = 1 if simulated_obs.shape[1] == 1: simulated_obs = simulated_obs[:, 0] return simulated_obs
[docs] def impulse_responses(self, params, steps=1, impulse=0, orthogonalized=False, cumulative=False, **kwargs): """ Impulse response function Parameters ---------- params : array_like Array of model parameters. steps : int, optional The number of steps for which impulse responses are calculated. Default is 1. Note that the initial impulse is not counted as a step, so if `steps=1`, the output will have 2 entries. impulse : int or array_like If an integer, the state innovation to pulse; must be between 0 and `k_posdef-1`. Alternatively, a custom impulse vector may be provided; must be shaped `k_posdef x 1`. orthogonalized : boolean, optional Whether or not to perform impulse using orthogonalized innovations. Note that this will also affect custum `impulse` vectors. Default is False. cumulative : boolean, optional Whether or not to return cumulative impulse responses. Default is False. **kwargs If the model is time-varying and `steps` is greater than the number of observations, any of the state space representation matrices that are time-varying must have updated values provided for the out-of-sample steps. For example, if `design` is a time-varying component, `nobs` is 10, and `steps` is 15, a (`k_endog` x `k_states` x 5) matrix must be provided with the new design matrix values. Returns ------- impulse_responses : array Responses for each endogenous variable due to the impulse given by the `impulse` argument. A (steps + 1 x k_endog) array. Notes ----- Intercepts in the measurement and state equation are ignored when calculating impulse responses. """ self.update(params) irfs = self.ssm.impulse_responses( steps, impulse, orthogonalized, cumulative, **kwargs) # IRF is (nobs x k_endog); don't want to squeeze in case of steps = 1 if irfs.shape[1] == 1: irfs = irfs[:, 0] return irfs
[docs] @classmethod def from_formula(cls, formula, data, subset=None): """ Not implemented for state space models """ raise NotImplementedError
[docs]class MLEResults(tsbase.TimeSeriesModelResults): r""" Class to hold results from fitting a state space model. Parameters ---------- model : MLEModel instance The fitted model instance params : array Fitted parameters filter_results : KalmanFilter instance The underlying state space model and Kalman filter output Attributes ---------- model : Model instance A reference to the model that was fit. filter_results : KalmanFilter instance The underlying state space model and Kalman filter output nobs : float The number of observations used to fit the model. params : array The parameters of the model. scale : float This is currently set to 1.0 unless the model uses concentrated filtering. See Also -------- MLEModel statsmodels.tsa.statespace.kalman_filter.FilterResults statsmodels.tsa.statespace.representation.FrozenRepresentation """ def __init__(self, model, params, results, cov_type='opg', cov_kwds=None, **kwargs): self.data = model.data scale = results.scale tsbase.TimeSeriesModelResults.__init__(self, model, params, normalized_cov_params=None, scale=scale) # Save the state space representation output self.filter_results = results if isinstance(results, SmootherResults): self.smoother_results = results else: self.smoother_results = None # Dimensions self.nobs = self.filter_results.nobs self.nobs_diffuse = self.filter_results.nobs_diffuse if self.nobs_diffuse > 0 and self.loglikelihood_burn > 0: warnings.warn('Care should be used when applying a loglikelihood' ' burn to a model with exact diffuse initialization.' ' Some results objects, e.g. degrees of freedom,' ' expect only one of the two to be set.') # This only excludes explicitly burned (usually approximate diffuse) # periods but does not exclude approximate diffuse periods. This is # because the loglikelihood remains valid for the initial periods in # the exact diffuse case (see DK, 2012, section 7.2) and so also do # e.g. information criteria (see DK, 2012, section 7.4) and the score # vector (see DK, 2012, section 7.3.3, equation 7.15). # However, other objects should be excluded in the diffuse periods # (e.g. the diffuse forecast errors, so in some cases a different # nobs_effective will have to be computed and used) self.nobs_effective = self.nobs - self.loglikelihood_burn P = self.filter_results.initial_diffuse_state_cov self.k_diffuse_states = 0 if P is None else np.sum(np.diagonal(P) == 1) # Degrees of freedom (see DK 2012, section 7.4) self.df_model = (self.params.size + self.k_diffuse_states + self.filter_results.filter_concentrated) self.df_resid = self.nobs_effective - self.df_model # Setup covariance matrix notes dictionary if not hasattr(self, 'cov_kwds'): self.cov_kwds = {} self.cov_type = cov_type # Setup the cache self._cache = resettable_cache() # Handle covariance matrix calculation if cov_kwds is None: cov_kwds = {} self._cov_approx_complex_step = ( cov_kwds.pop('approx_complex_step', True)) self._cov_approx_centered = cov_kwds.pop('approx_centered', False) try: self._rank = None self._get_robustcov_results(cov_type=cov_type, use_self=True, **cov_kwds) except np.linalg.LinAlgError: self._rank = 0 k_params = len(self.params) self.cov_params_default = np.zeros((k_params, k_params)) * np.nan self.cov_kwds['cov_type'] = ( 'Covariance matrix could not be calculated: singular.' ' information matrix.') self.model.update(self.params) # References of filter and smoother output extra_arrays = [ 'filtered_state', 'filtered_state_cov', 'predicted_state', 'predicted_state_cov', 'forecasts', 'forecasts_error', 'forecasts_error_cov', 'standardized_forecasts_error', 'forecasts_error_diffuse_cov', 'predicted_diffuse_state_cov', 'scaled_smoothed_estimator', 'scaled_smoothed_estimator_cov', 'smoothing_error', 'smoothed_state', 'smoothed_state_cov', 'smoothed_state_autocov', 'smoothed_measurement_disturbance', 'smoothed_state_disturbance', 'smoothed_measurement_disturbance_cov', 'smoothed_state_disturbance_cov'] for name in extra_arrays: setattr(self, name, getattr(self.filter_results, name, None)) # Handle removing data self._data_attr_model = getattr(self, '_data_attr_model', []) self._data_attr_model.extend(['ssm']) self._data_attr.extend(extra_arrays) self._data_attr.extend(['filter_results', 'smoother_results']) self.data_in_cache = getattr(self, 'data_in_cache', []) self.data_in_cache.extend([]) def _get_robustcov_results(self, cov_type='opg', **kwargs): """ Create new results instance with specified covariance estimator as default Note: creating new results instance currently not supported. Parameters ---------- cov_type : string the type of covariance matrix estimator to use. See Notes below kwargs : depends on cov_type Required or optional arguments for covariance calculation. See Notes below. Returns ------- results : results instance This method creates a new results instance with the requested covariance as the default covariance of the parameters. Inferential statistics like p-values and hypothesis tests will be based on this covariance matrix. Notes ----- The following covariance types and required or optional arguments are currently available: - 'opg' for the outer product of gradient estimator - 'oim' for the observed information matrix estimator, calculated using the method of Harvey (1989) - 'approx' for the observed information matrix estimator, calculated using a numerical approximation of the Hessian matrix. Uses complex step approximation by default, or uses finite differences if `approx_complex_step=False` in the `cov_kwds` dictionary. - 'robust' for an approximate (quasi-maximum likelihood) covariance matrix that may be valid even in the presense of some misspecifications. Intermediate calculations use the 'oim' method. - 'robust_approx' is the same as 'robust' except that the intermediate calculations use the 'approx' method. - 'none' for no covariance matrix calculation. """ use_self = kwargs.pop('use_self', False) if use_self: res = self else: raise NotImplementedError res = self.__class__( self.model, self.params, normalized_cov_params=self.normalized_cov_params, scale=self.scale) # Set the new covariance type res.cov_type = cov_type res.cov_kwds = {} # Calculate the new covariance matrix approx_complex_step = self._cov_approx_complex_step if approx_complex_step: approx_type_str = 'complex-step' elif self._cov_approx_centered: approx_type_str = 'centered finite differences' else: approx_type_str = 'finite differences' k_params = len(self.params) if k_params == 0: res.cov_params_default = np.zeros((0, 0)) res._rank = 0 res.cov_kwds['description'] = 'No parameters estimated.' elif cov_type == 'custom': res.cov_type = kwargs['custom_cov_type'] res.cov_params_default = kwargs['custom_cov_params'] res.cov_kwds['description'] = kwargs['custom_description'] res._rank = np.linalg.matrix_rank(res.cov_params_default) elif cov_type == 'none': res.cov_params_default = np.zeros((k_params, k_params)) * np.nan res._rank = np.nan res.cov_kwds['description'] = 'Covariance matrix not calculated.' elif self.cov_type == 'approx': res.cov_params_default = res.cov_params_approx res.cov_kwds['description'] = ( 'Covariance matrix calculated using numerical (%s)' ' differentiation.' % approx_type_str) elif self.cov_type == 'oim': res.cov_params_default = res.cov_params_oim res.cov_kwds['description'] = ( 'Covariance matrix calculated using the observed information' ' matrix (%s) described in Harvey (1989).' % approx_type_str) elif self.cov_type == 'opg': res.cov_params_default = res.cov_params_opg res.cov_kwds['description'] = ( 'Covariance matrix calculated using the outer product of' ' gradients (%s).' % approx_type_str) elif self.cov_type == 'robust' or self.cov_type == 'robust_oim': res.cov_params_default = res.cov_params_robust_oim res.cov_kwds['description'] = ( 'Quasi-maximum likelihood covariance matrix used for' ' robustness to some misspecifications; calculated using the' ' observed information matrix (%s) described in' ' Harvey (1989).' % approx_type_str) elif self.cov_type == 'robust_approx': res.cov_params_default = res.cov_params_robust_approx res.cov_kwds['description'] = ( 'Quasi-maximum likelihood covariance matrix used for' ' robustness to some misspecifications; calculated using' ' numerical (%s) differentiation.' % approx_type_str) else: raise NotImplementedError('Invalid covariance matrix type.') return res
[docs] @cache_readonly def aic(self): """ (float) Akaike Information Criterion """ # return -2 * self.llf + 2 * self.df_model return aic(self.llf, self.nobs_effective, self.df_model)
[docs] @cache_readonly def bic(self): """ (float) Bayes Information Criterion """ # return (-2 * self.llf + # self.df_model * np.log(self.nobs_effective)) return bic(self.llf, self.nobs_effective, self.df_model)
def _cov_params_approx(self, approx_complex_step=True, approx_centered=False): evaluated_hessian = self.nobs_effective * self.model.hessian( params=self.params, transformed=True, method='approx', approx_complex_step=approx_complex_step, approx_centered=approx_centered) # TODO: Case with "not approx_complex_step" is not hit in # tests as of 2017-05-19 (neg_cov, singular_values) = pinv_extended(evaluated_hessian) self.model.update(self.params) if self._rank is None: self._rank = np.linalg.matrix_rank(np.diag(singular_values)) return -neg_cov
[docs] @cache_readonly def cov_params_approx(self): """ (array) The variance / covariance matrix. Computed using the numerical Hessian approximated by complex step or finite differences methods. """ return self._cov_params_approx(self._cov_approx_complex_step, self._cov_approx_centered)
def _cov_params_oim(self, approx_complex_step=True, approx_centered=False): evaluated_hessian = self.nobs_effective * self.model.hessian( self.params, hessian_method='oim', transformed=True, approx_complex_step=approx_complex_step, approx_centered=approx_centered) (neg_cov, singular_values) = pinv_extended(evaluated_hessian) self.model.update(self.params) if self._rank is None: self._rank = np.linalg.matrix_rank(np.diag(singular_values)) return -neg_cov
[docs] @cache_readonly def cov_params_oim(self): """ (array) The variance / covariance matrix. Computed using the method from Harvey (1989). """ return self._cov_params_oim(self._cov_approx_complex_step, self._cov_approx_centered)
def _cov_params_opg(self, approx_complex_step=True, approx_centered=False): evaluated_hessian = self.nobs_effective * self.model._hessian_opg( self.params, transformed=True, approx_complex_step=approx_complex_step, approx_centered=approx_centered) (neg_cov, singular_values) = pinv_extended(evaluated_hessian) self.model.update(self.params) if self._rank is None: self._rank = np.linalg.matrix_rank(np.diag(singular_values)) return -neg_cov
[docs] @cache_readonly def cov_params_opg(self): """ (array) The variance / covariance matrix. Computed using the outer product of gradients method. """ return self._cov_params_opg(self._cov_approx_complex_step, self._cov_approx_centered)
[docs] @cache_readonly def cov_params_robust(self): """ (array) The QMLE variance / covariance matrix. Alias for `cov_params_robust_oim` """ return self.cov_params_robust_oim
def _cov_params_robust_oim(self, approx_complex_step=True, approx_centered=False): cov_opg = self._cov_params_opg(approx_complex_step=approx_complex_step, approx_centered=approx_centered) evaluated_hessian = self.nobs_effective * self.model.hessian( self.params, hessian_method='oim', transformed=True, approx_complex_step=approx_complex_step, approx_centered=approx_centered) cov_params, singular_values = pinv_extended( np.dot(np.dot(evaluated_hessian, cov_opg), evaluated_hessian) ) self.model.update(self.params) if self._rank is None: self._rank = np.linalg.matrix_rank(np.diag(singular_values)) return cov_params
[docs] @cache_readonly def cov_params_robust_oim(self): """ (array) The QMLE variance / covariance matrix. Computed using the method from Harvey (1989) as the evaluated hessian. """ return self._cov_params_robust_oim(self._cov_approx_complex_step, self._cov_approx_centered)
def _cov_params_robust_approx(self, approx_complex_step=True, approx_centered=False): cov_opg = self._cov_params_opg(approx_complex_step=approx_complex_step, approx_centered=approx_centered) evaluated_hessian = self.nobs_effective * self.model.hessian( self.params, transformed=True, method='approx', approx_complex_step=approx_complex_step) # TODO: Case with "not approx_complex_step" is not # hit in tests as of 2017-05-19 (cov_params, singular_values) = pinv_extended( np.dot(np.dot(evaluated_hessian, cov_opg), evaluated_hessian) ) self.model.update(self.params) if self._rank is None: self._rank = np.linalg.matrix_rank(np.diag(singular_values)) return cov_params
[docs] @cache_readonly def cov_params_robust_approx(self): """ (array) The QMLE variance / covariance matrix. Computed using the numerical Hessian as the evaluated hessian. """ return self._cov_params_robust_approx(self._cov_approx_complex_step, self._cov_approx_centered)
[docs] def info_criteria(self, criteria, method='standard'): r""" Information criteria Parameters ---------- criteria : {'aic', 'bic', 'hqic'} The information criteria to compute. method : {'standard', 'lutkepohl'} The method for information criteria computation. Default is 'standard' method; 'lutkepohl' computes the information criteria as in Lütkepohl (2007). See Notes for formulas. Notes ----- The `'standard'` formulas are: .. math:: AIC & = -2 \log L(Y_n | \hat \psi) + 2 k \\ BIC & = -2 \log L(Y_n | \hat \psi) + k \log n \\ HQIC & = -2 \log L(Y_n | \hat \psi) + 2 k \log \log n \\ where :math:`\hat \psi` are the maximum likelihood estimates of the parameters, :math:`n` is the number of observations, and `k` is the number of estimated parameters. Note that the `'standard'` formulas are returned from the `aic`, `bic`, and `hqic` results attributes. The `'lutkepohl'` formuals are (Lütkepohl, 2010): .. math:: AIC_L & = \log | Q | + \frac{2 k}{n} \\ BIC_L & = \log | Q | + \frac{k \log n}{n} \\ HQIC_L & = \log | Q | + \frac{2 k \log \log n}{n} \\ where :math:`Q` is the state covariance matrix. Note that the Lütkepohl definitions do not apply to all state space models, and should be used with care outside of SARIMAX and VARMAX models. References ---------- .. [*] Lütkepohl, Helmut. 2007. *New Introduction to Multiple Time* *Series Analysis.* Berlin: Springer. """ criteria = criteria.lower() method = method.lower() if method == 'standard': out = getattr(self, criteria) elif method == 'lutkepohl': if self.filter_results.state_cov.shape[-1] > 1: raise ValueError('Cannot compute Lütkepohl statistics for' ' models with time-varying state covariance' ' matrix.') cov = self.filter_results.state_cov[:, :, 0] if criteria == 'aic': out = np.squeeze(np.linalg.slogdet(cov)[1] + 2 * self.df_model / self.nobs_effective) elif criteria == 'bic': out = np.squeeze(np.linalg.slogdet(cov)[1] + self.df_model * np.log(self.nobs_effective) / self.nobs_effective) elif criteria == 'hqic': out = np.squeeze(np.linalg.slogdet(cov)[1] + 2 * self.df_model * np.log(np.log(self.nobs_effective)) / self.nobs_effective) else: raise ValueError('Invalid information criteria') else: raise ValueError('Invalid information criteria computation method') return out
[docs] @cache_readonly def fittedvalues(self): """ (array) The predicted values of the model. An (nobs x k_endog) array. """ # This is a (k_endog x nobs array; don't want to squeeze in case of # the corner case where nobs = 1 (mostly a concern in the predict or # forecast functions, but here also to maintain consistency) fittedvalues = self.forecasts if fittedvalues.shape[0] == 1: fittedvalues = fittedvalues[0, :] else: fittedvalues = fittedvalues.T return fittedvalues
[docs] @cache_readonly def hqic(self): """ (float) Hannan-Quinn Information Criterion """ # return (-2 * self.llf + # 2 * np.log(np.log(self.nobs_effective)) * self.df_model) return hqic(self.llf, self.nobs_effective, self.df_model)
[docs] @cache_readonly def llf_obs(self): """ (float) The value of the log-likelihood function evaluated at `params`. """ return self.filter_results.llf_obs
[docs] @cache_readonly def llf(self): """ (float) The value of the log-likelihood function evaluated at `params`. """ return self.llf_obs[self.filter_results.loglikelihood_burn:].sum()
[docs] @cache_readonly def loglikelihood_burn(self): """ (float) The number of observations during which the likelihood is not evaluated. """ return self.filter_results.loglikelihood_burn
[docs] @cache_readonly def pvalues(self): """ (array) The p-values associated with the z-statistics of the coefficients. Note that the coefficients are assumed to have a Normal distribution. """ return norm.sf(np.abs(self.zvalues)) * 2
[docs] @cache_readonly def resid(self): """ (array) The model residuals. An (nobs x k_endog) array. """ # This is a (k_endog x nobs array; don't want to squeeze in case of # the corner case where nobs = 1 (mostly a concern in the predict or # forecast functions, but here also to maintain consistency) resid = self.forecasts_error if resid.shape[0] == 1: resid = resid[0, :] else: resid = resid.T return resid
[docs] @cache_readonly def zvalues(self): """ (array) The z-statistics for the coefficients. """ return self.params / self.bse
[docs] def test_normality(self, method): """ Test for normality of standardized residuals. Null hypothesis is normality. Parameters ---------- method : string {'jarquebera'} or None The statistical test for normality. Must be 'jarquebera' for Jarque-Bera normality test. If None, an attempt is made to select an appropriate test. Notes ----- Let `d` = max(loglikelihood_burn, nobs_diffuse); this test is calculated ignoring the first `d` residuals. In the case of missing data, the maintained hypothesis is that the data are missing completely at random. This test is then run on the standardized residuals excluding those corresponding to missing observations. See Also -------- statsmodels.stats.stattools.jarque_bera """ if method is None: method = 'jarquebera' if method == 'jarquebera': from statsmodels.stats.stattools import jarque_bera d = np.maximum(self.loglikelihood_burn, self.nobs_diffuse) output = [] for i in range(self.model.k_endog): resid = self.filter_results.standardized_forecasts_error[i, d:] mask = ~np.isnan(resid) output.append(jarque_bera(resid[mask])) else: raise NotImplementedError('Invalid normality test method.') return np.array(output)
[docs] def test_heteroskedasticity(self, method, alternative='two-sided', use_f=True): r""" Test for heteroskedasticity of standardized residuals Tests whether the sum-of-squares in the first third of the sample is significantly different than the sum-of-squares in the last third of the sample. Analogous to a Goldfeld-Quandt test. The null hypothesis is of no heteroskedasticity. Parameters ---------- method : string {'breakvar'} or None The statistical test for heteroskedasticity. Must be 'breakvar' for test of a break in the variance. If None, an attempt is made to select an appropriate test. alternative : string, 'increasing', 'decreasing' or 'two-sided' This specifies the alternative for the p-value calculation. Default is two-sided. use_f : boolean, optional Whether or not to compare against the asymptotic distribution (chi-squared) or the approximate small-sample distribution (F). Default is True (i.e. default is to compare against an F distribution). Returns ------- output : array An array with `(test_statistic, pvalue)` for each endogenous variable. The array is then sized `(k_endog, 2)`. If the method is called as `het = res.test_heteroskedasticity()`, then `het[0]` is an array of size 2 corresponding to the first endogenous variable, where `het[0][0]` is the test statistic, and `het[0][1]` is the p-value. Notes ----- The null hypothesis is of no heteroskedasticity. That means different things depending on which alternative is selected: - Increasing: Null hypothesis is that the variance is not increasing throughout the sample; that the sum-of-squares in the later subsample is *not* greater than the sum-of-squares in the earlier subsample. - Decreasing: Null hypothesis is that the variance is not decreasing throughout the sample; that the sum-of-squares in the earlier subsample is *not* greater than the sum-of-squares in the later subsample. - Two-sided: Null hypothesis is that the variance is not changing throughout the sample. Both that the sum-of-squares in the earlier subsample is not greater than the sum-of-squares in the later subsample *and* that the sum-of-squares in the later subsample is not greater than the sum-of-squares in the earlier subsample. For :math:`h = [T/3]`, the test statistic is: .. math:: H(h) = \sum_{t=T-h+1}^T \tilde v_t^2 \Bigg / \sum_{t=d+1}^{d+1+h} \tilde v_t^2 where :math:`d` = max(loglikelihood_burn, nobs_diffuse)` (usually corresponding to diffuse initialization under either the approximate or exact approach). This statistic can be tested against an :math:`F(h,h)` distribution. Alternatively, :math:`h H(h)` is asymptotically distributed according to :math:`\chi_h^2`; this second test can be applied by passing `asymptotic=True` as an argument. See section 5.4 of [1]_ for the above formula and discussion, as well as additional details. TODO - Allow specification of :math:`h` References ---------- .. [1] Harvey, Andrew C. 1990. *Forecasting, Structural Time Series* *Models and the Kalman Filter.* Cambridge University Press. """ if method is None: method = 'breakvar' if method == 'breakvar': # Store some values squared_resid = self.filter_results.standardized_forecasts_error**2 d = np.maximum(self.loglikelihood_burn, self.nobs_diffuse) # This differs from self.nobs_effective because here we want to # exclude exact diffuse periods, whereas self.nobs_effective only # excludes explicitly burned (usually approximate diffuse) periods. nobs_effective = self.nobs - d test_statistics = [] p_values = [] for i in range(self.model.k_endog): h = int(np.round(nobs_effective / 3)) numer_resid = squared_resid[i, -h:] numer_resid = numer_resid[~np.isnan(numer_resid)] numer_dof = len(numer_resid) denom_resid = squared_resid[i, d:d+h] denom_resid = denom_resid[~np.isnan(denom_resid)] denom_dof = len(denom_resid) if numer_dof < 2: warnings.warn('Early subset of data for variable %d' ' has too few non-missing observations to' ' calculate test statistic.' % i) numer_resid = np.nan if denom_dof < 2: warnings.warn('Later subset of data for variable %d' ' has too few non-missing observations to' ' calculate test statistic.' % i) denom_resid = np.nan test_statistic = np.sum(numer_resid) / np.sum(denom_resid) # Setup functions to calculate the p-values if use_f: from scipy.stats import f pval_lower = lambda test_statistics: f.cdf( # noqa:E731 test_statistics, numer_dof, denom_dof) pval_upper = lambda test_statistics: f.sf( # noqa:E731 test_statistics, numer_dof, denom_dof) else: from scipy.stats import chi2 pval_lower = lambda test_statistics: chi2.cdf( # noqa:E731 numer_dof * test_statistics, denom_dof) pval_upper = lambda test_statistics: chi2.sf( # noqa:E731 numer_dof * test_statistics, denom_dof) # Calculate the one- or two-sided p-values alternative = alternative.lower() if alternative in ['i', 'inc', 'increasing']: p_value = pval_upper(test_statistic) elif alternative in ['d', 'dec', 'decreasing']: test_statistic = 1. / test_statistic p_value = pval_upper(test_statistic) elif alternative in ['2', '2-sided', 'two-sided']: p_value = 2 * np.minimum( pval_lower(test_statistic), pval_upper(test_statistic) ) else: raise ValueError('Invalid alternative.') test_statistics.append(test_statistic) p_values.append(p_value) output = np.c_[test_statistics, p_values] else: raise NotImplementedError('Invalid heteroskedasticity test' ' method.') return output
[docs] def test_serial_correlation(self, method, lags=None): """ Ljung-box test for no serial correlation of standardized residuals Null hypothesis is no serial correlation. Parameters ---------- method : string {'ljungbox','boxpierece'} or None The statistical test for serial correlation. If None, an attempt is made to select an appropriate test. lags : None, int or array_like If lags is an integer then this is taken to be the largest lag that is included, the test result is reported for all smaller lag length. If lags is a list or array, then all lags are included up to the largest lag in the list, however only the tests for the lags in the list are reported. If lags is None, then the default maxlag is 12*(nobs/100)^{1/4} Returns ------- output : array An array with `(test_statistic, pvalue)` for each endogenous variable and each lag. The array is then sized `(k_endog, 2, lags)`. If the method is called as `ljungbox = res.test_serial_correlation()`, then `ljungbox[i]` holds the results of the Ljung-Box test (as would be returned by `statsmodels.stats.diagnostic.acorr_ljungbox`) for the `i` th endogenous variable. Notes ----- Let `d` = max(loglikelihood_burn, nobs_diffuse); this test is calculated ignoring the first `d` residuals. Output is nan for any endogenous variable which has missing values. See Also -------- statsmodels.stats.diagnostic.acorr_ljungbox """ if method is None: method = 'ljungbox' if method == 'ljungbox' or method == 'boxpierce': from statsmodels.stats.diagnostic import acorr_ljungbox d = np.maximum(self.loglikelihood_burn, self.nobs_diffuse) # This differs from self.nobs_effective because here we want to # exclude exact diffuse periods, whereas self.nobs_effective only # excludes explicitly burned (usually approximate diffuse) periods. nobs_effective = self.nobs - d output = [] # Default lags for acorr_ljungbox is 40, but may not always have # that many observations if lags is None: lags = min(40, nobs_effective - 1) for i in range(self.model.k_endog): results = acorr_ljungbox( self.filter_results.standardized_forecasts_error[i][d:], lags=lags, boxpierce=(method == 'boxpierce')) if method == 'ljungbox': output.append(results[0:2]) else: output.append(results[2:]) output = np.c_[output] else: raise NotImplementedError('Invalid serial correlation test' ' method.') return output
[docs] def get_prediction(self, start=None, end=None, dynamic=False, index=None, **kwargs): """ In-sample prediction and out-of-sample forecasting Parameters ---------- start : int, str, or datetime, optional Zero-indexed observation number at which to start forecasting, i.e., the first forecast is start. Can also be a date string to parse or a datetime type. Default is the the zeroth observation. end : int, str, or datetime, optional Zero-indexed observation number at which to end forecasting, i.e., the last forecast is end. Can also be a date string to parse or a datetime type. However, if the dates index does not have a fixed frequency, end must be an integer index if you want out of sample prediction. Default is the last observation in the sample. dynamic : boolean, int, str, or datetime, optional Integer offset relative to `start` at which to begin dynamic prediction. Can also be an absolute date string to parse or a datetime type (these are not interpreted as offsets). Prior to this observation, true endogenous values will be used for prediction; starting with this observation and continuing through the end of prediction, forecasted endogenous values will be used instead. **kwargs Additional arguments may required for forecasting beyond the end of the sample. See `FilterResults.predict` for more details. Returns ------- forecast : array Array of out of in-sample predictions and / or out-of-sample forecasts. An (npredict x k_endog) array. """ if start is None: start = self.model._index[0] # Handle start, end, dynamic start, end, out_of_sample, prediction_index = ( self.model._get_prediction_index(start, end, index)) # Handle `dynamic` if isinstance(dynamic, (bytes, unicode)): dynamic, _, _ = self.model._get_index_loc(dynamic) # Perform the prediction # This is a (k_endog x npredictions) array; don't want to squeeze in # case of npredictions = 1 prediction_results = self.filter_results.predict( start, end + out_of_sample + 1, dynamic, **kwargs) # Return a new mlemodel.PredictionResults object return PredictionResultsWrapper(PredictionResults( self, prediction_results, row_labels=prediction_index))
[docs] def get_forecast(self, steps=1, **kwargs): """ Out-of-sample forecasts Parameters ---------- steps : int, str, or datetime, optional If an integer, the number of steps to forecast from the end of the sample. Can also be a date string to parse or a datetime type. However, if the dates index does not have a fixed frequency, steps must be an integer. Default **kwargs Additional arguments may required for forecasting beyond the end of the sample. See `FilterResults.predict` for more details. Returns ------- forecast : array Array of out of sample forecasts. A (steps x k_endog) array. """ if isinstance(steps, (int, long)): end = self.nobs + steps - 1 else: end = steps return self.get_prediction(start=self.nobs, end=end, **kwargs)
[docs] def predict(self, start=None, end=None, dynamic=False, **kwargs): """ In-sample prediction and out-of-sample forecasting Parameters ---------- start : int, str, or datetime, optional Zero-indexed observation number at which to start forecasting, i.e., the first forecast is start. Can also be a date string to parse or a datetime type. Default is the the zeroth observation. end : int, str, or datetime, optional Zero-indexed observation number at which to end forecasting, i.e., the last forecast is end. Can also be a date string to parse or a datetime type. However, if the dates index does not have a fixed frequency, end must be an integer index if you want out of sample prediction. Default is the last observation in the sample. dynamic : boolean, int, str, or datetime, optional Integer offset relative to `start` at which to begin dynamic prediction. Can also be an absolute date string to parse or a datetime type (these are not interpreted as offsets). Prior to this observation, true endogenous values will be used for prediction; starting with this observation and continuing through the end of prediction, forecasted endogenous values will be used instead. **kwargs Additional arguments may required for forecasting beyond the end of the sample. See `FilterResults.predict` for more details. Returns ------- forecast : array Array of out of in-sample predictions and / or out-of-sample forecasts. An (npredict x k_endog) array. """ # Perform the prediction prediction_results = self.get_prediction(start, end, dynamic, **kwargs) return prediction_results.predicted_mean
[docs] def forecast(self, steps=1, **kwargs): """ Out-of-sample forecasts Parameters ---------- steps : int, str, or datetime, optional If an integer, the number of steps to forecast from the end of the sample. Can also be a date string to parse or a datetime type. However, if the dates index does not have a fixed frequency, steps must be an integer. Default **kwargs Additional arguments may required for forecasting beyond the end of the sample. See `FilterResults.predict` for more details. Returns ------- forecast : array Array of out of sample forecasts. A (steps x k_endog) array. """ if isinstance(steps, (int, long)): end = self.nobs + steps - 1 else: end = steps return self.predict(start=self.nobs, end=end, **kwargs)
[docs] def simulate(self, nsimulations, measurement_shocks=None, state_shocks=None, initial_state=None): r""" Simulate a new time series following the state space model Parameters ---------- nsimulations : int The number of observations to simulate. If the model is time-invariant this can be any number. If the model is time-varying, then this number must be less than or equal to the number measurement_shocks : array_like, optional If specified, these are the shocks to the measurement equation, :math:`\varepsilon_t`. If unspecified, these are automatically generated using a pseudo-random number generator. If specified, must be shaped `nsimulations` x `k_endog`, where `k_endog` is the same as in the state space model. state_shocks : array_like, optional If specified, these are the shocks to the state equation, :math:`\eta_t`. If unspecified, these are automatically generated using a pseudo-random number generator. If specified, must be shaped `nsimulations` x `k_posdef` where `k_posdef` is the same as in the state space model. initial_state : array_like, optional If specified, this is the state vector at time zero, which should be shaped (`k_states` x 1), where `k_states` is the same as in the state space model. If unspecified, but the model has been initialized, then that initialization is used. If unspecified and the model has not been initialized, then a vector of zeros is used. Note that this is not included in the returned `simulated_states` array. Returns ------- simulated_obs : array An (nsimulations x k_endog) array of simulated observations. """ scale = self.scale if self.filter_results.filter_concentrated else None with self.model.ssm.fixed_scale(scale): sim = self.model.simulate(self.params, nsimulations, measurement_shocks, state_shocks, initial_state) return sim
[docs] def impulse_responses(self, steps=1, impulse=0, orthogonalized=False, cumulative=False, **kwargs): """ Impulse response function Parameters ---------- steps : int, optional The number of steps for which impulse responses are calculated. Default is 1. Note that the initial impulse is not counted as a step, so if `steps=1`, the output will have 2 entries. impulse : int or array_like If an integer, the state innovation to pulse; must be between 0 and `k_posdef-1`. Alternatively, a custom impulse vector may be provided; must be shaped `k_posdef x 1`. orthogonalized : boolean, optional Whether or not to perform impulse using orthogonalized innovations. Note that this will also affect custum `impulse` vectors. Default is False. cumulative : boolean, optional Whether or not to return cumulative impulse responses. Default is False. **kwargs If the model is time-varying and `steps` is greater than the number of observations, any of the state space representation matrices that are time-varying must have updated values provided for the out-of-sample steps. For example, if `design` is a time-varying component, `nobs` is 10, and `steps` is 15, a (`k_endog` x `k_states` x 5) matrix must be provided with the new design matrix values. Returns ------- impulse_responses : array Responses for each endogenous variable due to the impulse given by the `impulse` argument. A (steps + 1 x k_endog) array. Notes ----- Intercepts in the measurement and state equation are ignored when calculating impulse responses. """ scale = self.scale if self.filter_results.filter_concentrated else None with self.model.ssm.fixed_scale(scale): irfs = self.model.impulse_responses(self.params, steps, impulse, orthogonalized, cumulative, **kwargs) return irfs
[docs] def plot_diagnostics(self, variable=0, lags=10, fig=None, figsize=None): """ Diagnostic plots for standardized residuals of one endogenous variable Parameters ---------- variable : integer, optional Index of the endogenous variable for which the diagnostic plots should be created. Default is 0. lags : integer, optional Number of lags to include in the correlogram. Default is 10. fig : Matplotlib Figure instance, optional If given, subplots are created in this figure instead of in a new figure. Note that the 2x2 grid will be created in the provided figure using `fig.add_subplot()`. figsize : tuple, optional If a figure is created, this argument allows specifying a size. The tuple is (width, height). Notes ----- Produces a 2x2 plot grid with the following plots (ordered clockwise from top left): 1. Standardized residuals over time 2. Histogram plus estimated density of standardized residulas, along with a Normal(0,1) density plotted for reference. 3. Normal Q-Q plot, with Normal reference line. 4. Correlogram See Also -------- statsmodels.graphics.gofplots.qqplot statsmodels.graphics.tsaplots.plot_acf """ from statsmodels.graphics.utils import _import_mpl, create_mpl_fig _import_mpl() fig = create_mpl_fig(fig, figsize) # Eliminate residuals associated with burned or diffuse likelihoods d = np.maximum(self.loglikelihood_burn, self.nobs_diffuse) resid = self.filter_results.standardized_forecasts_error[variable, d:] # Top-left: residuals vs time ax = fig.add_subplot(221) if hasattr(self.data, 'dates') and self.data.dates is not None: x = self.data.dates[d:]._mpl_repr() else: x = np.arange(len(resid)) ax.plot(x, resid) ax.hlines(0, x[0], x[-1], alpha=0.5) ax.set_xlim(x[0], x[-1]) ax.set_title('Standardized residual') # Top-right: histogram, Gaussian kernel density, Normal density # Can only do histogram and Gaussian kernel density on the non-null # elements resid_nonmissing = resid[~(np.isnan(resid))] ax = fig.add_subplot(222) # temporarily disable Deprecation warning, normed -> density # hist needs to use `density` in future when minimum matplotlib has it with warnings.catch_warnings(record=True): ax.hist(resid_nonmissing, normed=True, label='Hist') from scipy.stats import gaussian_kde, norm kde = gaussian_kde(resid_nonmissing) xlim = (-1.96*2, 1.96*2) x = np.linspace(xlim[0], xlim[1]) ax.plot(x, kde(x), label='KDE') ax.plot(x, norm.pdf(x), label='N(0,1)') ax.set_xlim(xlim) ax.legend() ax.set_title('Histogram plus estimated density') # Bottom-left: QQ plot ax = fig.add_subplot(223) from statsmodels.graphics.gofplots import qqplot qqplot(resid_nonmissing, line='s', ax=ax) ax.set_title('Normal Q-Q') # Bottom-right: Correlogram ax = fig.add_subplot(224) from statsmodels.graphics.tsaplots import plot_acf plot_acf(resid, ax=ax, lags=lags) ax.set_title('Correlogram') ax.set_ylim(-1, 1) return fig
[docs] def summary(self, alpha=.05, start=None, title=None, model_name=None, display_params=True): """ Summarize the Model Parameters ---------- alpha : float, optional Significance level for the confidence intervals. Default is 0.05. start : int, optional Integer of the start observation. Default is 0. model_name : string The name of the model used. Default is to use model class name. Returns ------- summary : Summary instance This holds the summary table and text, which can be printed or converted to various output formats. See Also -------- statsmodels.iolib.summary.Summary """ from statsmodels.iolib.summary import Summary # Model specification results model = self.model if title is None: title = 'Statespace Model Results' if start is None: start = 0 if self.model._index_dates: ix = self.model._index d = ix[start] sample = ['%02d-%02d-%02d' % (d.month, d.day, d.year)] d = ix[-1] sample += ['- ' + '%02d-%02d-%02d' % (d.month, d.day, d.year)] else: sample = [str(start), ' - ' + str(self.nobs)] # Standardize the model name as a list of str if model_name is None: model_name = model.__class__.__name__ # Diagnostic tests results try: het = self.test_heteroskedasticity(method='breakvar') except Exception: # FIXME: catch something specific het = np.array([[np.nan]*2]) try: lb = self.test_serial_correlation(method='ljungbox') except Exception: # FIXME: catch something specific lb = np.array([[np.nan]*2]).reshape(1, 2, 1) try: jb = self.test_normality(method='jarquebera') except Exception: # FIXME: catch something specific jb = np.array([[np.nan]*4]) # Create the tables if not isinstance(model_name, list): model_name = [model_name] top_left = [('Dep. Variable:', None)] top_left.append(('Model:', [model_name[0]])) for i in range(1, len(model_name)): top_left.append(('', ['+ ' + model_name[i]])) top_left += [ ('Date:', None), ('Time:', None), ('Sample:', [sample[0]]), ('', [sample[1]]) ] top_right = [ ('No. Observations:', [self.nobs]), ('Log Likelihood', ["%#5.3f" % self.llf]), ] if hasattr(self, 'rsquared'): top_right.append(('R-squared:', ["%#8.3f" % self.rsquared])) top_right += [ ('AIC', ["%#5.3f" % self.aic]), ('BIC', ["%#5.3f" % self.bic]), ('HQIC', ["%#5.3f" % self.hqic])] if (self.filter_results is not None and self.filter_results.filter_concentrated): top_right.append(('Scale', ["%#5.3f" % self.scale])) if hasattr(self, 'cov_type'): top_left.append(('Covariance Type:', [self.cov_type])) format_str = lambda array: [ # noqa:E731 ', '.join(['{0:.2f}'.format(i) for i in array]) ] diagn_left = [('Ljung-Box (Q):', format_str(lb[:, 0, -1])), ('Prob(Q):', format_str(lb[:, 1, -1])), ('Heteroskedasticity (H):', format_str(het[:, 0])), ('Prob(H) (two-sided):', format_str(het[:, 1])) ] diagn_right = [('Jarque-Bera (JB):', format_str(jb[:, 0])), ('Prob(JB):', format_str(jb[:, 1])), ('Skew:', format_str(jb[:, 2])), ('Kurtosis:', format_str(jb[:, 3])) ] summary = Summary() summary.add_table_2cols(self, gleft=top_left, gright=top_right, title=title) if len(self.params) > 0 and display_params: summary.add_table_params(self, alpha=alpha, xname=self.data.param_names, use_t=False) summary.add_table_2cols(self, gleft=diagn_left, gright=diagn_right, title="") # Add warnings/notes, added to text format only etext = [] if hasattr(self, 'cov_type') and 'description' in self.cov_kwds: etext.append(self.cov_kwds['description']) if self._rank < len(self.params): etext.append("Covariance matrix is singular or near-singular," " with condition number %6.3g. Standard errors may be" " unstable." % np.linalg.cond(self.cov_params())) if etext: etext = ["[{0}] {1}".format(i + 1, text) for i, text in enumerate(etext)] etext.insert(0, "Warnings:") summary.add_extra_txt(etext) return summary
class MLEResultsWrapper(wrap.ResultsWrapper): _attrs = { 'zvalues': 'columns', 'cov_params_approx': 'cov', 'cov_params_default': 'cov', 'cov_params_oim': 'cov', 'cov_params_opg': 'cov', 'cov_params_robust': 'cov', 'cov_params_robust_approx': 'cov', 'cov_params_robust_oim': 'cov', } _wrap_attrs = wrap.union_dicts(tsbase.TimeSeriesResultsWrapper._wrap_attrs, _attrs) _methods = { 'forecast': 'dates', 'simulate': 'ynames', 'impulse_responses': 'ynames' } _wrap_methods = wrap.union_dicts( tsbase.TimeSeriesResultsWrapper._wrap_methods, _methods) wrap.populate_wrapper(MLEResultsWrapper, MLEResults) # noqa:E305 class PredictionResults(pred.PredictionResults): """ Parameters ---------- prediction_results : kalman_filter.PredictionResults instance Results object from prediction after fitting or filtering a state space model. row_labels : iterable Row labels for the predicted data. Attributes ---------- """ def __init__(self, model, prediction_results, row_labels=None): if model.model.k_endog == 1: endog = pd.Series(prediction_results.endog[:, 0], name=model.model.endog_names) else: endog = pd.DataFrame(prediction_results.endog.T, columns=model.model.endog_names) self.model = Bunch(data=model.data.__class__( endog=endog, predict_dates=row_labels)) self.prediction_results = prediction_results # Get required values predicted_mean = self.prediction_results.forecasts if predicted_mean.shape[0] == 1: predicted_mean = predicted_mean[0, :] else: predicted_mean = predicted_mean.transpose() var_pred_mean = self.prediction_results.forecasts_error_cov if var_pred_mean.shape[0] == 1: var_pred_mean = var_pred_mean[0, 0, :] else: var_pred_mean = var_pred_mean.transpose() # Initialize super(PredictionResults, self).__init__(predicted_mean, var_pred_mean, dist='norm', row_labels=row_labels, link=identity()) @property def se_mean(self): if self.var_pred_mean.ndim == 1: se_mean = np.sqrt(self.var_pred_mean) else: se_mean = np.sqrt(self.var_pred_mean.T.diagonal()) return se_mean def conf_int(self, method='endpoint', alpha=0.05, **kwds): # TODO: this performs metadata wrapping, and that should be handled # by attach_* methods. However, they don't currently support # this use case. conf_int = super(PredictionResults, self).conf_int( method, alpha, **kwds) # Create a dataframe if self.row_labels is not None: conf_int = pd.DataFrame(conf_int, index=self.row_labels) # Attach the endog names ynames = self.model.data.ynames if not type(ynames) == list: ynames = [ynames] names = (['lower %s' % name for name in ynames] + ['upper %s' % name for name in ynames]) conf_int.columns = names return conf_int def summary_frame(self, endog=0, what='all', alpha=0.05): # TODO: finish and cleanup # import pandas as pd from collections import OrderedDict # ci_obs = self.conf_int(alpha=alpha, obs=True) # need to split ci_mean = np.asarray(self.conf_int(alpha=alpha)) to_include = OrderedDict() if self.predicted_mean.ndim == 1: yname = self.model.data.ynames to_include['mean'] = self.predicted_mean to_include['mean_se'] = self.se_mean k_endog = 1 else: yname = self.model.data.ynames[endog] to_include['mean'] = self.predicted_mean[:, endog] to_include['mean_se'] = self.se_mean[:, endog] k_endog = self.predicted_mean.shape[1] to_include['mean_ci_lower'] = ci_mean[:, endog] to_include['mean_ci_upper'] = ci_mean[:, k_endog + endog] # OrderedDict doesn't work to preserve sequence # pandas dict doesn't handle 2d_array # data = np.column_stack(list(to_include.values())) # names = .... res = pd.DataFrame(to_include, index=self.row_labels, columns=to_include.keys()) res.columns.name = yname return res class PredictionResultsWrapper(wrap.ResultsWrapper): _attrs = { 'predicted_mean': 'dates', 'se_mean': 'dates', 't_values': 'dates', } _wrap_attrs = wrap.union_dicts(_attrs) _methods = {} _wrap_methods = wrap.union_dicts(_methods) wrap.populate_wrapper(PredictionResultsWrapper, PredictionResults) # noqa:E305