""" State Space Representation and Kalman Filter, Smoother Author: Chad Fulton License: Simplified-BSD """ import numpy as np from types import SimpleNamespace from statsmodels.tsa.statespace.representation import OptionWrapper from statsmodels.tsa.statespace.kalman_filter import (KalmanFilter, FilterResults) from statsmodels.tsa.statespace.tools import ( reorder_missing_matrix, reorder_missing_vector, copy_index_matrix) from statsmodels.tsa.statespace import tools, initialization SMOOTHER_STATE = 0x01 # Durbin and Koopman (2012), Chapter 4.4.2 SMOOTHER_STATE_COV = 0x02 # ibid., Chapter 4.4.3 SMOOTHER_DISTURBANCE = 0x04 # ibid., Chapter 4.5 SMOOTHER_DISTURBANCE_COV = 0x08 # ibid., Chapter 4.5 SMOOTHER_STATE_AUTOCOV = 0x10 # ibid., Chapter 4.7 SMOOTHER_ALL = ( SMOOTHER_STATE | SMOOTHER_STATE_COV | SMOOTHER_DISTURBANCE | SMOOTHER_DISTURBANCE_COV | SMOOTHER_STATE_AUTOCOV ) SMOOTH_CONVENTIONAL = 0x01 SMOOTH_CLASSICAL = 0x02 SMOOTH_ALTERNATIVE = 0x04 SMOOTH_UNIVARIATE = 0x08 class KalmanSmoother(KalmanFilter): r""" State space representation of a time series process, with Kalman filter and smoother. Parameters ---------- k_endog : {array_like, int} The observed time-series process :math:`y` if array like or the number of variables in the process if an integer. k_states : int The dimension of the unobserved state process. k_posdef : int, optional The dimension of a guaranteed positive definite covariance matrix describing the shocks in the measurement equation. Must be less than or equal to `k_states`. Default is `k_states`. results_class : class, optional Default results class to use to save filtering output. Default is `SmootherResults`. If specified, class must extend from `SmootherResults`. **kwargs Keyword arguments may be used to provide default values for state space matrices, for Kalman filtering options, or for Kalman smoothing options. See `Representation` for more details. """ smoother_outputs = [ 'smoother_state', 'smoother_state_cov', 'smoother_state_autocov', 'smoother_disturbance', 'smoother_disturbance_cov', 'smoother_all', ] smoother_state = OptionWrapper('smoother_output', SMOOTHER_STATE) smoother_state_cov = OptionWrapper('smoother_output', SMOOTHER_STATE_COV) smoother_disturbance = ( OptionWrapper('smoother_output', SMOOTHER_DISTURBANCE) ) smoother_disturbance_cov = ( OptionWrapper('smoother_output', SMOOTHER_DISTURBANCE_COV) ) smoother_state_autocov = ( OptionWrapper('smoother_output', SMOOTHER_STATE_AUTOCOV) ) smoother_all = OptionWrapper('smoother_output', SMOOTHER_ALL) smooth_methods = [ 'smooth_conventional', 'smooth_alternative', 'smooth_classical' ] smooth_conventional = OptionWrapper('smooth_method', SMOOTH_CONVENTIONAL) """ (bool) Flag for conventional (Durbin and Koopman, 2012) Kalman smoothing. """ smooth_alternative = OptionWrapper('smooth_method', SMOOTH_ALTERNATIVE) """ (bool) Flag for alternative (modified Bryson-Frazier) smoothing. """ smooth_classical = OptionWrapper('smooth_method', SMOOTH_CLASSICAL) """ (bool) Flag for classical (see e.g. Anderson and Moore, 1979) smoothing. """ smooth_univariate = OptionWrapper('smooth_method', SMOOTH_UNIVARIATE) """ (bool) Flag for univariate smoothing (uses modified Bryson-Frazier timing). """ # Default smoother options smoother_output = SMOOTHER_ALL smooth_method = 0 def __init__(self, k_endog, k_states, k_posdef=None, results_class=None, kalman_smoother_classes=None, **kwargs): # Set the default results class if results_class is None: results_class = SmootherResults # Extract keyword arguments to-be-used later keys = ['smoother_output'] + KalmanSmoother.smoother_outputs smoother_output_kwargs = {key: kwargs.pop(key) for key in keys if key in kwargs} keys = ['smooth_method'] + KalmanSmoother.smooth_methods smooth_method_kwargs = {key: kwargs.pop(key) for key in keys if key in kwargs} # Initialize the base class super().__init__( k_endog, k_states, k_posdef, results_class=results_class, **kwargs ) # Options self.prefix_kalman_smoother_map = ( kalman_smoother_classes if kalman_smoother_classes is not None else tools.prefix_kalman_smoother_map.copy()) # Setup the underlying Kalman smoother storage self._kalman_smoothers = {} # Set the smoother options self.set_smoother_output(**smoother_output_kwargs) self.set_smooth_method(**smooth_method_kwargs) def _clone_kwargs(self, endog, **kwargs): # See Representation._clone_kwargs for docstring kwargs = super()._clone_kwargs(endog, **kwargs) # Get defaults for options kwargs.setdefault('smoother_output', self.smoother_output) kwargs.setdefault('smooth_method', self.smooth_method) return kwargs @property def _kalman_smoother(self): prefix = self.prefix if prefix in self._kalman_smoothers: return self._kalman_smoothers[prefix] return None def _initialize_smoother(self, smoother_output=None, smooth_method=None, prefix=None, **kwargs): if smoother_output is None: smoother_output = self.smoother_output if smooth_method is None: smooth_method = self.smooth_method # Make sure we have the required Kalman filter prefix, dtype, create_filter, create_statespace = ( self._initialize_filter(prefix, **kwargs) ) # Determine if we need to (re-)create the smoother # (definitely need to recreate if we recreated the filter) create_smoother = (create_filter or prefix not in self._kalman_smoothers) if not create_smoother: kalman_smoother = self._kalman_smoothers[prefix] create_smoother = (kalman_smoother.kfilter is not self._kalman_filters[prefix]) # If the dtype-specific _kalman_smoother does not exist (or if we # need to re-create it), create it if create_smoother: # Setup the smoother cls = self.prefix_kalman_smoother_map[prefix] self._kalman_smoothers[prefix] = cls( self._statespaces[prefix], self._kalman_filters[prefix], smoother_output, smooth_method ) # Otherwise, update the smoother parameters else: self._kalman_smoothers[prefix].set_smoother_output( smoother_output, False) self._kalman_smoothers[prefix].set_smooth_method(smooth_method) return prefix, dtype, create_smoother, create_filter, create_statespace 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 : int, 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. See notes for details. Notes ----- The smoother output is defined by a collection of boolean flags, and is internally stored as a bitmask. The methods available are: SMOOTHER_STATE = 0x01 Calculate and return the smoothed states. SMOOTHER_STATE_COV = 0x02 Calculate and return the smoothed state covariance matrices. SMOOTHER_STATE_AUTOCOV = 0x10 Calculate and return the smoothed state lag-one autocovariance matrices. SMOOTHER_DISTURBANCE = 0x04 Calculate and return the smoothed state and observation disturbances. SMOOTHER_DISTURBANCE_COV = 0x08 Calculate and return the covariance matrices for the smoothed state and observation disturbances. SMOOTHER_ALL Calculate and return all results. If the bitmask is set directly via the `smoother_output` argument, then the full method must be provided. If keyword arguments are used to set individual boolean flags, then the lowercase of the method must be used as an argument name, and the value is the desired value of the boolean flag (True or False). Note that the smoother output may also be specified by directly modifying the class attributes which are defined similarly to the keyword arguments. The default smoother output is SMOOTHER_ALL. If performance is a concern, only those results which are needed should be specified as any results that are not specified will not be calculated. For example, if the smoother output is set to only include SMOOTHER_STATE, the smoother operates much more quickly than if all output is required. Examples -------- >>> import statsmodels.tsa.statespace.kalman_smoother as ks >>> mod = ks.KalmanSmoother(1,1) >>> mod.smoother_output 15 >>> mod.set_smoother_output(smoother_output=0) >>> mod.smoother_state = True >>> mod.smoother_output 1 >>> mod.smoother_state True """ if smoother_output is not None: self.smoother_output = smoother_output for name in KalmanSmoother.smoother_outputs: if name in kwargs: setattr(self, name, kwargs[name]) def set_smooth_method(self, smooth_method=None, **kwargs): r""" Set the smoothing method The smoothing method can be used to override the Kalman smoother approach used. By default, the Kalman smoother used depends on the Kalman filter method. Parameters ---------- smooth_method : int, 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 ----- The smoothing method is defined by a collection of boolean flags, and is internally stored as a bitmask. The methods available are: SMOOTH_CONVENTIONAL = 0x01 Default Kalman smoother, as presented in Durbin and Koopman, 2012 chapter 4. SMOOTH_CLASSICAL = 0x02 Classical Kalman smoother, as presented in Anderson and Moore, 1979 or Durbin and Koopman, 2012 chapter 4.6.1. SMOOTH_ALTERNATIVE = 0x04 Modified Bryson-Frazier Kalman smoother method; this is identical to the conventional method of Durbin and Koopman, 2012, except that an additional intermediate step is included. SMOOTH_UNIVARIATE = 0x08 Univariate Kalman smoother, as presented in Durbin and Koopman, 2012 chapter 6, except with modified Bryson-Frazier timing. Practically speaking, these methods should all produce the same output but different computational implications, numerical stability implications, or internal timing assumptions. Note that only the first method is available if using a Scipy version older than 0.16. If the bitmask is set directly via the `smooth_method` argument, then the full method must be provided. If keyword arguments are used to set individual boolean flags, then the lowercase of the method must be used as an argument name, and the value is the desired value of the boolean flag (True or False). Note that the filter method may also be specified by directly modifying the class attributes which are defined similarly to the keyword arguments. The default filtering method is SMOOTH_CONVENTIONAL. Examples -------- >>> mod = sm.tsa.statespace.SARIMAX(range(10)) >>> mod.smooth_method 1 >>> mod.filter_conventional True >>> mod.filter_univariate = True >>> mod.smooth_method 17 >>> mod.set_smooth_method(filter_univariate=False, filter_collapsed=True) >>> mod.smooth_method 33 >>> mod.set_smooth_method(smooth_method=1) >>> mod.filter_conventional True >>> mod.filter_univariate False >>> mod.filter_collapsed False >>> mod.filter_univariate = True >>> mod.smooth_method 17 """ if smooth_method is not None: self.smooth_method = smooth_method for name in KalmanSmoother.smooth_methods: if name in kwargs: setattr(self, name, kwargs[name]) def _smooth(self, smoother_output=None, smooth_method=None, prefix=None, complex_step=False, results=None, **kwargs): # Initialize the smoother prefix, dtype, create_smoother, create_filter, create_statespace = ( self._initialize_smoother( smoother_output, smooth_method, prefix=prefix, **kwargs )) # Check that the filter and statespace weren't just recreated if create_filter or create_statespace: raise ValueError('Passed settings forced re-creation of the' ' Kalman filter. Please run `_filter` before' ' running `_smooth`.') # Get the appropriate smoother smoother = self._kalman_smoothers[prefix] # Run the smoother smoother() return smoother def smooth(self, smoother_output=None, smooth_method=None, results=None, run_filter=True, prefix=None, complex_step=False, update_representation=True, update_filter=True, update_smoother=True, **kwargs): """ Apply the Kalman smoother to the statespace model. Parameters ---------- smoother_output : int, optional Determines which Kalman smoother output calculate. Default is all (including state, disturbances, and all covariances). results : class or object, optional If a class, then that class is instantiated and returned with the result of both filtering and smoothing. If an object, then that object is updated with the smoothing data. If None, then a SmootherResults object is returned with both filtering and smoothing results. run_filter : bool, optional Whether or not to run the Kalman filter prior to smoothing. Default is True. prefix : str The prefix of the datatype. Usually only used internally. Returns ------- SmootherResults object """ # Run the filter kfilter = self._filter(**kwargs) # Create the results object results = self.results_class(self) if update_representation: results.update_representation(self) if update_filter: results.update_filter(kfilter) else: # (even if we don't update all filter results, still need to # update this) results.nobs_diffuse = kfilter.nobs_diffuse # Run the smoother if smoother_output is None: smoother_output = self.smoother_output smoother = self._smooth(smoother_output, results=results, **kwargs) # Update the results if update_smoother: results.update_smoother(smoother) return results class SmootherResults(FilterResults): r""" Results from applying the Kalman smoother and/or filter to a state space model. Parameters ---------- model : Representation A Statespace representation Attributes ---------- nobs : int Number of observations. k_endog : int The dimension of the observation series. k_states : int The dimension of the unobserved state process. k_posdef : int The dimension of a guaranteed positive definite covariance matrix describing the shocks in the measurement equation. dtype : dtype Datatype of representation matrices prefix : str BLAS prefix of representation matrices shapes : dictionary of name:tuple A dictionary recording the shapes of each of the representation matrices as tuples. endog : ndarray The observation vector. design : ndarray The design matrix, :math:`Z`. obs_intercept : ndarray The intercept for the observation equation, :math:`d`. obs_cov : ndarray The covariance matrix for the observation equation :math:`H`. transition : ndarray The transition matrix, :math:`T`. state_intercept : ndarray The intercept for the transition equation, :math:`c`. selection : ndarray The selection matrix, :math:`R`. state_cov : ndarray The covariance matrix for the state equation :math:`Q`. missing : array of bool An array of the same size as `endog`, filled with boolean values that are True if the corresponding entry in `endog` is NaN and False otherwise. nmissing : array of int An array of size `nobs`, where the ith entry is the number (between 0 and k_endog) of NaNs in the ith row of the `endog` array. time_invariant : bool Whether or not the representation matrices are time-invariant initialization : str Kalman filter initialization method. initial_state : array_like The state vector used to initialize the Kalamn filter. initial_state_cov : array_like The state covariance matrix used to initialize the Kalamn filter. filter_method : int Bitmask representing the Kalman filtering method inversion_method : int Bitmask representing the method used to invert the forecast error covariance matrix. stability_method : int Bitmask representing the methods used to promote numerical stability in the Kalman filter recursions. conserve_memory : int Bitmask representing the selected memory conservation method. tolerance : float The tolerance at which the Kalman filter determines convergence to steady-state. loglikelihood_burn : int The number of initial periods during which the loglikelihood is not recorded. converged : bool Whether or not the Kalman filter converged. period_converged : int The time period in which the Kalman filter converged. filtered_state : ndarray The filtered state vector at each time period. filtered_state_cov : ndarray The filtered state covariance matrix at each time period. predicted_state : ndarray The predicted state vector at each time period. predicted_state_cov : ndarray The predicted state covariance matrix at each time period. kalman_gain : ndarray The Kalman gain at each time period. forecasts : ndarray The one-step-ahead forecasts of observations at each time period. forecasts_error : ndarray The forecast errors at each time period. forecasts_error_cov : ndarray The forecast error covariance matrices at each time period. loglikelihood : ndarray The loglikelihood values at each time period. collapsed_forecasts : ndarray If filtering using collapsed observations, stores the one-step-ahead forecasts of collapsed observations at each time period. collapsed_forecasts_error : ndarray If filtering using collapsed observations, stores the one-step-ahead forecast errors of collapsed observations at each time period. collapsed_forecasts_error_cov : ndarray If filtering using collapsed observations, stores the one-step-ahead forecast error covariance matrices of collapsed observations at each time period. standardized_forecast_error : ndarray The standardized forecast errors smoother_output : int Bitmask representing the generated Kalman smoothing output scaled_smoothed_estimator : ndarray The scaled smoothed estimator at each time period. scaled_smoothed_estimator_cov : ndarray The scaled smoothed estimator covariance matrices at each time period. smoothing_error : ndarray The smoothing error covariance matrices at each time period. smoothed_state : ndarray The smoothed state at each time period. smoothed_state_cov : ndarray The smoothed state covariance matrices at each time period. smoothed_state_autocov : ndarray The smoothed state lago-one autocovariance matrices at each time period: :math:`Cov(\alpha_{t+1}, \alpha_t)`. smoothed_measurement_disturbance : ndarray The smoothed measurement at each time period. smoothed_state_disturbance : ndarray The smoothed state at each time period. smoothed_measurement_disturbance_cov : ndarray The smoothed measurement disturbance covariance matrices at each time period. smoothed_state_disturbance_cov : ndarray The smoothed state disturbance covariance matrices at each time period. """ _smoother_attributes = [ 'smoother_output', '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', 'innovations_transition' ] _smoother_options = KalmanSmoother.smoother_outputs _attributes = FilterResults._model_attributes + _smoother_attributes def update_representation(self, model, only_options=False): """ Update the results to match a given model Parameters ---------- model : Representation The model object from which to take the updated values. only_options : bool, optional If set to true, only the smoother and filter options are updated, and the state space representation is not updated. Default is False. Notes ----- This method is rarely required except for internal usage. """ super().update_representation(model, only_options) # Save the options as boolean variables for name in self._smoother_options: setattr(self, name, getattr(model, name, None)) # Initialize holders for smoothed forecasts self._smoothed_forecasts = None self._smoothed_forecasts_error = None self._smoothed_forecasts_error_cov = None def update_smoother(self, smoother): """ Update the smoother results Parameters ---------- smoother : KalmanSmoother The model object from which to take the updated values. Notes ----- This method is rarely required except for internal usage. """ # Copy the appropriate output attributes = [] # Since update_representation will already have been called, we can # use the boolean options smoother_* and know they match the smoother # itself if self.smoother_state or self.smoother_disturbance: attributes.append('scaled_smoothed_estimator') if self.smoother_state_cov or self.smoother_disturbance_cov: attributes.append('scaled_smoothed_estimator_cov') if self.smoother_state: attributes.append('smoothed_state') if self.smoother_state_cov: attributes.append('smoothed_state_cov') if self.smoother_state_autocov: attributes.append('smoothed_state_autocov') if self.smoother_disturbance: attributes += [ 'smoothing_error', 'smoothed_measurement_disturbance', 'smoothed_state_disturbance' ] if self.smoother_disturbance_cov: attributes += [ 'smoothed_measurement_disturbance_cov', 'smoothed_state_disturbance_cov' ] has_missing = np.sum(self.nmissing) > 0 for name in self._smoother_attributes: if name == 'smoother_output': pass elif name in attributes: if name in ['smoothing_error', 'smoothed_measurement_disturbance']: vector = getattr(smoother, name, None) if vector is not None and has_missing: vector = np.array(reorder_missing_vector( vector, self.missing, prefix=self.prefix)) else: vector = np.array(vector, copy=True) setattr(self, name, vector) elif name == 'smoothed_measurement_disturbance_cov': matrix = getattr(smoother, name, None) if matrix is not None and has_missing: matrix = reorder_missing_matrix( matrix, self.missing, reorder_rows=True, reorder_cols=True, prefix=self.prefix) # In the missing data case, we want to set the missing # components equal to their unconditional distribution copy_index_matrix( self.obs_cov, matrix, self.missing, index_rows=True, index_cols=True, inplace=True, prefix=self.prefix) else: matrix = np.array(matrix, copy=True) setattr(self, name, matrix) else: setattr(self, name, np.array(getattr(smoother, name, None), copy=True)) else: setattr(self, name, None) self.innovations_transition = ( np.array(smoother.innovations_transition, copy=True)) # Diffuse objects self.scaled_smoothed_diffuse_estimator = None self.scaled_smoothed_diffuse1_estimator_cov = None self.scaled_smoothed_diffuse2_estimator_cov = None if self.nobs_diffuse > 0: self.scaled_smoothed_diffuse_estimator = np.array( smoother.scaled_smoothed_diffuse_estimator, copy=True) self.scaled_smoothed_diffuse1_estimator_cov = np.array( smoother.scaled_smoothed_diffuse1_estimator_cov, copy=True) self.scaled_smoothed_diffuse2_estimator_cov = np.array( smoother.scaled_smoothed_diffuse2_estimator_cov, copy=True) # Adjustments # For r_t (and similarly for N_t), what was calculated was # r_T, ..., r_{-1}. We only want r_0, ..., r_T # so exclude the appropriate element so that the time index is # consistent with the other returned output # r_t stored such that scaled_smoothed_estimator[0] == r_{-1} start = 1 end = None if 'scaled_smoothed_estimator' in attributes: self.scaled_smoothed_estimator_presample = ( self.scaled_smoothed_estimator[:, 0]) self.scaled_smoothed_estimator = ( self.scaled_smoothed_estimator[:, start:end] ) if 'scaled_smoothed_estimator_cov' in attributes: self.scaled_smoothed_estimator_cov_presample = ( self.scaled_smoothed_estimator_cov[:, :, 0]) self.scaled_smoothed_estimator_cov = ( self.scaled_smoothed_estimator_cov[:, :, start:end] ) # Clear the smoothed forecasts self._smoothed_forecasts = None self._smoothed_forecasts_error = None self._smoothed_forecasts_error_cov = None # Note: if we concentrated out the scale, need to adjust the # loglikelihood values and all of the covariance matrices and the # values that depend on the covariance matrices if self.filter_concentrated and self.model._scale is None: self.smoothed_state_cov *= self.scale self.smoothed_state_autocov *= self.scale self.smoothed_state_disturbance_cov *= self.scale self.smoothed_measurement_disturbance_cov *= self.scale self.scaled_smoothed_estimator_presample /= self.scale self.scaled_smoothed_estimator /= self.scale self.scaled_smoothed_estimator_cov_presample /= self.scale self.scaled_smoothed_estimator_cov /= self.scale self.smoothing_error /= self.scale # Cache self.__smoothed_state_autocovariance = {} def _smoothed_state_autocovariance(self, shift, start, end, extend_kwargs=None): """ Compute "forward" autocovariances, Cov(t, t+j) Parameters ---------- shift : int The number of period to shift forwards when computing the autocovariance. This has the opposite sign as `lag` from the `smoothed_state_autocovariance` method. start : int, optional The start of the interval (inclusive) of autocovariances to compute and return. end : int, optional The end of the interval (exclusive) autocovariances to compute and return. Note that since it is an exclusive endpoint, the returned autocovariances do not include the value at this index. extend_kwargs : dict, optional Keyword arguments containing updated state space system matrices for handling out-of-sample autocovariance computations in time-varying state space models. """ if extend_kwargs is None: extend_kwargs = {} # Size of returned array in the time dimension n = end - start # Get number of post-sample periods we need to create an extended # model to compute if shift == 0: max_insample = self.nobs - shift else: max_insample = self.nobs - shift + 1 n_postsample = max(0, end - max_insample) # Get full in-sample arrays if shift != 0: L = self.innovations_transition P = self.predicted_state_cov N = self.scaled_smoothed_estimator_cov else: acov = self.smoothed_state_cov # If applicable, append out-of-sample arrays if n_postsample > 0: # Note: we need 1 less than the number of post endog = np.zeros((n_postsample, self.k_endog)) * np.nan mod = self.model.extend(endog, start=self.nobs, **extend_kwargs) mod.initialize_known(self.predicted_state[..., self.nobs], self.predicted_state_cov[..., self.nobs]) res = mod.smooth() if shift != 0: start_insample = max(0, start) L = np.concatenate((L[..., start_insample:], res.innovations_transition), axis=2) P = np.concatenate((P[..., start_insample:], res.predicted_state_cov[..., 1:]), axis=2) N = np.concatenate((N[..., start_insample:], res.scaled_smoothed_estimator_cov), axis=2) end -= start_insample start -= start_insample else: acov = np.concatenate((acov, res.predicted_state_cov), axis=2) if shift != 0: # Subset to appropriate start, end start_insample = max(0, start) LT = L[..., start_insample:end + shift - 1].T P = P[..., start_insample:end + shift].T N = N[..., start_insample:end + shift - 1].T # Intermediate computations tmpLT = np.eye(self.k_states)[None, :, :] length = P.shape[0] - shift # this is the required length of LT for i in range(1, shift + 1): tmpLT = LT[shift - i:length + shift - i] @ tmpLT eye = np.eye(self.k_states)[None, ...] # Compute the autocovariance acov = np.zeros((n, self.k_states, self.k_states)) acov[:start_insample - start] = np.nan acov[start_insample - start:] = ( P[:-shift] @ tmpLT @ (eye - N[shift - 1:] @ P[shift:])) else: acov = acov.T[start:end] return acov def smoothed_state_autocovariance(self, lag=1, t=None, start=None, end=None, extend_kwargs=None): r""" Compute state vector autocovariances, conditional on the full dataset Computes: .. math:: Cov(\alpha_t - \hat \alpha_t, \alpha_{t - j} - \hat \alpha_{t - j}) where the `lag` argument gives the value for :math:`j`. Thus when the `lag` argument is positive, the autocovariance is between the current and previous periods, while if `lag` is negative the autocovariance is between the current and future periods. Parameters ---------- lag : int, optional The number of period to shift when computing the autocovariance. Default is 1. t : int, optional A specific period for which to compute and return the autocovariance. Cannot be used in combination with `start` or `end`. See the Returns section for details on how this parameter affects what is what is returned. start : int, optional The start of the interval (inclusive) of autocovariances to compute and return. Cannot be used in combination with the `t` argument. See the Returns section for details on how this parameter affects what is what is returned. Default is 0. end : int, optional The end of the interval (exclusive) autocovariances to compute and return. Note that since it is an exclusive endpoint, the returned autocovariances do not include the value at this index. Cannot be used in combination with the `t` argument. See the Returns section for details on how this parameter affects what is what is returned and what the default value is. extend_kwargs : dict, optional Keyword arguments containing updated state space system matrices for handling out-of-sample autocovariance computations in time-varying state space models. Returns ------- acov : ndarray Array of autocovariance matrices. If the argument `t` is not provided, then it is shaped `(k_states, k_states, n)`, while if `t` given then the third axis is dropped and the array is shaped `(k_states, k_states)`. The output under the default case differs somewhat based on the state space model and the sign of the lag. To see how these cases differ, denote the output at each time point as Cov(t, t-j). Then: - If `lag > 0` (and the model is either time-varying or time-invariant), then the returned array is shaped `(*, *, nobs)` and each entry [:, :, t] contains Cov(t, t-j). However, the model does not have enough information to compute autocovariances in the pre-sample period, so that we cannot compute Cov(1, 1-lag), Cov(2, 2-lag), ..., Cov(lag, 0). Thus the first `lag` entries have all values set to NaN. - If the model is time-invariant and `lag < -1` or if `lag` is 0 or -1, and the model is either time-invariant or time-varying, then the returned array is shaped `(*, *, nobs)` and each entry [:, :, t] contains Cov(t, t+j). Moreover, all entries are available (i.e. there are no NaNs). - If the model is time-varying and `lag < -1` and `extend_kwargs` is not provided, then the returned array is shaped `(*, *, nobs - lag + 1)`. - However, if the model is time-varying and `lag < -1`, then `extend_kwargs` can be provided with `lag - 1` additional matrices so that the returned array is shaped `(*, *, nobs)` as usual. More generally, the dimension of the last axis will be `start - end`. Notes ----- This method computes: .. math:: Cov(\alpha_t - \hat \alpha_t, \alpha_{t - j} - \hat \alpha_{t - j}) where the `lag` argument determines the autocovariance order :math:`j`, and `lag` is an integer (positive, zero, or negative). This method cannot compute values associated with time points prior to the sample, and so it returns a matrix of NaN values for these time points. For example, if `start=0` and `lag=2`, then assuming the output is assigned to the variable `acov`, we will have `acov[..., 0]` and `acov[..., 1]` as matrices filled with NaN values. Based only on the "current" results object (i.e. the Kalman smoother applied to the sample), there is not enough information to compute Cov(t, t+j) for the last `lag - 1` observations of the sample. However, the values can be computed for these time points using the transition equation of the state space representation, and so for time-invariant state space models we do compute these values. For time-varying models, this can also be done, but updated state space matrices for the out-of-sample time points must be provided via the `extend_kwargs` argument. See [1]_, Chapter 4.7, for all details about how these autocovariances are computed. The `t` and `start`/`end` parameters compute and return only the requested autocovariances. As a result, using these parameters is recommended to reduce the computational burden, particularly if the number of observations and/or the dimension of the state vector is large. References ---------- .. [1] Durbin, James, and Siem Jan Koopman. 2012. Time Series Analysis by State Space Methods: Second Edition. Oxford University Press. """ # We can cache the results for time-invariant models cache_key = None if extend_kwargs is None or len(extend_kwargs) == 0: cache_key = (lag, t, start, end) # Short-circuit for a cache-hit if (cache_key is not None and cache_key in self.__smoothed_state_autocovariance): return self.__smoothed_state_autocovariance[cache_key] # Switch to only positive values for `lag` forward_autocovariances = False if lag < 0: lag = -lag forward_autocovariances = True # Handle `t` if t is not None and (start is not None or end is not None): raise ValueError('Cannot specify both `t` and `start` or `end`.') if t is not None: start = t end = t + 1 # Defaults if start is None: start = 0 if end is None: if forward_autocovariances and lag > 1 and extend_kwargs is None: end = self.nobs - lag + 1 else: end = self.nobs if extend_kwargs is None: extend_kwargs = {} # Sanity checks if start < 0 or end < 0: raise ValueError('Negative `t`, `start`, or `end` is not allowed.') if end < start: raise ValueError('`end` must be after `start`') if lag == 0 and self.smoothed_state_cov is None: raise RuntimeError('Cannot return smoothed state covariances' ' if those values have not been computed by' ' Kalman smoothing.') # We already have in-sample (+1 out-of-sample) smoothed covariances if lag == 0 and end <= self.nobs + 1: acov = self.smoothed_state_cov if end == self.nobs + 1: acov = np.concatenate( (acov[..., start:], self.predicted_state_cov[..., -1:]), axis=2).T else: acov = acov.T[start:end] # In-sample, we can compute up to Cov(T, T+1) or Cov(T+1, T) and down # to Cov(1, 2) or Cov(2, 1). So: # - For lag=1 we set Cov(1, 0) = np.nan and then can compute up to T-1 # in-sample values Cov(2, 1), ..., Cov(T, T-1) and the first # out-of-sample value Cov(T+1, T) elif (lag == 1 and self.smoothed_state_autocov is not None and not forward_autocovariances and end <= self.nobs + 1): # nans = np.zeros((self.k_states, self.k_states, lag)) * np.nan # acov = np.concatenate((nans, self.smoothed_state_autocov), # axis=2).transpose(2, 0, 1)[start:end] if start == 0: nans = np.zeros((self.k_states, self.k_states, lag)) * np.nan acov = np.concatenate( (nans, self.smoothed_state_autocov[..., :end - 1]), axis=2) else: acov = self.smoothed_state_autocov[..., start - 1:end - 1] acov = acov.transpose(2, 0, 1) # - For lag=-1 we can compute T in-sample values, Cov(1, 2), ..., # Cov(T, T+1) but we cannot compute the first out-of-sample value # Cov(T+1, T+2). elif (lag == 1 and self.smoothed_state_autocov is not None and forward_autocovariances and end < self.nobs + 1): acov = self.smoothed_state_autocov.T[start:end] # Otherwise, we need to compute additional values at the end of the # sample else: if forward_autocovariances: # Cov(t, t + lag), t = start, ..., end acov = self._smoothed_state_autocovariance( lag, start, end, extend_kwargs=extend_kwargs) else: # Cov(t, t + lag)' = Cov(t + lag, t), # with t = start - lag, ..., end - lag out = self._smoothed_state_autocovariance( lag, start - lag, end - lag, extend_kwargs=extend_kwargs) acov = out.transpose(0, 2, 1) # Squeeze the last axis or else reshape to have the same axis # definitions as e.g. smoothed_state_cov if t is not None: acov = acov[0] else: acov = acov.transpose(1, 2, 0) # Fill in the cache, if applicable if cache_key is not None: self.__smoothed_state_autocovariance[cache_key] = acov return acov def news(self, previous, t=None, start=None, end=None, revisions_details_start=True, design=None, state_index=None): r""" Compute the news and impacts associated with a data release Parameters ---------- previous : SmootherResults Prior results object relative to which to compute the news. This results object must have identical state space representation for the prior sample period so that the only difference is that this results object has updates to the observed data. t : int, optional A specific period for which to compute the news. Cannot be used in combination with `start` or `end`. start : int, optional The start of the interval (inclusive) of news to compute. Cannot be used in combination with the `t` argument. Default is the last period of the sample (`nobs - 1`). end : int, optional The end of the interval (exclusive) of news to compute. Note that since it is an exclusive endpoint, the returned news do not include the value at this index. Cannot be used in combination with the `t` argument. revisions_details_start : bool or int, optional The period at which to beging computing the detailed impacts of data revisions. Any revisions prior to this period will have their impacts grouped together. If a negative integer, interpreted as an offset from the end of the dataset. If set to True, detailed impacts are computed for all revisions, while if set to False, all revisions are grouped together. Default is False. Note that for large models, setting this to be near the beginning of the sample can cause this function to be slow. design : array, optional Design matrix for the period `t` in time-varying models. If this model has a time-varying design matrix, and the argument `t` is out of this model's sample, then a new design matrix for period `t` must be provided. Unused otherwise. state_index : array_like, optional An optional index specifying a subset of states to use when constructing the impacts of revisions and news. For example, if `state_index=[0, 1]` is passed, then only the impacts to the observed variables arising from the impacts to the first two states will be returned. Returns ------- news_results : SimpleNamespace News and impacts associated with a data release. Includes the following attributes: - `update_impacts`: update to forecasts of impacted variables from the news. It is equivalent to E[y^i | post] - E[y^i | revision], where y^i are the variables of interest. In [1]_, this is described as "revision" in equation (17). - `revision_detailed_impacts`: update to forecasts of variables impacted variables from data revisions. It is E[y^i | revision] - E[y^i | previous], and does not have a specific notation in [1]_, since there for simplicity they assume that there are no revisions. - `news`: the unexpected component of the updated data. Denoted I = y^u - E[y^u | previous], where y^u are the data points that were newly incorporated in a data release (but not including revisions to data points that already existed in the previous release). In [1]_, this is described as "news" in equation (17). - `revisions`: y^r(updated) - y^r(previous) for periods in which detailed impacts were computed - `revisions_all` : y^r(updated) - y^r(previous) for all revisions - `gain`: the gain matrix associated with the "Kalman-like" update from the news, E[y I'] E[I I']^{-1}. In [1]_, this can be found in the equation For E[y_{k,t_k} \mid I_{v+1}] in the middle of page 17. - `revision_weights` weights on observations for the smoothed signal - `update_forecasts`: forecasts of the updated periods used to construct the news, E[y^u | previous]. - `update_realized`: realizations of the updated periods used to construct the news, y^u. - `revised`: revised observations of the periods that were revised and for which detailed impacts were computed - `revised`: revised observations of the periods that were revised - `revised_prev`: previous observations of the periods that were revised and for which detailed impacts were computed - `revised_prev_all`: previous observations of the periods that were revised and for which detailed impacts were computed - `prev_impacted_forecasts`: previous forecast of the periods of interest, E[y^i | previous]. - `post_impacted_forecasts`: forecast of the periods of interest after taking into account both revisions and updates, E[y^i | post]. - `revision_results`: results object that updates the `previous` results to take into account data revisions. - `revision_results`: results object associated with the revisions - `revision_impacts`: total impacts from all revisions (both grouped and detailed) - `revisions_ix`: list of `(t, i)` positions of revisions in endog - `revisions_details`: list of `(t, i)` positions of revisions to endog for which details of impacts were computed - `revisions_grouped`: list of `(t, i)` positions of revisions to endog for which impacts were grouped - `revisions_details_start`: period in which revision details start to be computed - `updates_ix`: list of `(t, i)` positions of updates to endog - `state_index`: index of state variables used to compute impacts Notes ----- This method computes the effect of new data (e.g. from a new data release) on smoothed forecasts produced by a state space model, as described in [1]_. It also computes the effect of revised data on smoothed forecasts. References ---------- .. [1] Bańbura, Marta and Modugno, Michele. 2010. "Maximum likelihood estimation of factor models on data sets with arbitrary pattern of missing data." No 1189, Working Paper Series, European Central Bank. https://EconPapers.repec.org/RePEc:ecb:ecbwps:20101189. .. [2] Bańbura, Marta, and Michele Modugno. "Maximum likelihood estimation of factor models on datasets with arbitrary pattern of missing data." Journal of Applied Econometrics 29, no. 1 (2014): 133-160. """ # Handle `t` if t is not None and (start is not None or end is not None): raise ValueError('Cannot specify both `t` and `start` or `end`.') if t is not None: start = t end = t + 1 # Defaults if start is None: start = self.nobs - 1 if end is None: end = self.nobs # Sanity checks if start < 0 or end < 0: raise ValueError('Negative `t`, `start`, or `end` is not allowed.') if end <= start: raise ValueError('`end` must be after `start`') if self.smoothed_state_cov is None: raise ValueError('Cannot compute news without having applied the' ' Kalman smoother first.') error_ss = ('This results object has %s and so it does not appear to' ' by an extension of `previous`. Can only compute the' ' news by comparing this results set to previous results' ' objects.') if self.nobs < previous.nobs: raise ValueError(error_ss % 'fewer observations than' ' `previous`') if not (self.k_endog == previous.k_endog and self.k_states == previous.k_states and self.k_posdef == previous.k_posdef): raise ValueError(error_ss % 'different state space dimensions than' ' `previous`') for key in self.model.shapes.keys(): if key == 'obs': continue tv = getattr(self, key).shape[-1] > 1 tv_prev = getattr(previous, key).shape[-1] > 1 if tv and not tv_prev: raise ValueError(error_ss % f'time-varying {key} while' ' `previous` does not') if not tv and tv_prev: raise ValueError(error_ss % f'time-invariant {key} while' ' `previous` does not') # Standardize if state_index is not None: state_index = np.atleast_1d( np.sort(np.array(state_index, dtype=int))) # We cannot forecast out-of-sample periods in a time-varying model if end > self.nobs and not self.model.time_invariant: raise RuntimeError('Cannot compute the impacts of news on periods' ' outside of the sample in time-varying' ' models.') # For time-varying case, figure out extension kwargs extend_kwargs = {} for key in self.model.shapes.keys(): if key == 'obs': continue mat = getattr(self, key) prev_mat = getattr(previous, key) if mat.shape[-1] > prev_mat.shape[-1]: extend_kwargs[key] = mat[..., prev_mat.shape[-1]:] # Figure out which indices have changed revisions_ix, updates_ix = previous.model.diff_endog(self.endog.T) # Compute prev / post impact forecasts prev_impacted_forecasts = previous.predict( start=start, end=end, **extend_kwargs).smoothed_forecasts post_impacted_forecasts = self.predict( start=start, end=end).smoothed_forecasts # Separate revisions into those with detailed impacts and those where # impacts are grouped together if revisions_details_start is True: revisions_details_start = 0 elif revisions_details_start is False: revisions_details_start = previous.nobs elif revisions_details_start < 0: revisions_details_start = previous.nobs + revisions_details_start revisions_grouped = [] revisions_details = [] if revisions_details_start > 0: for s, i in revisions_ix: if s < revisions_details_start: revisions_grouped.append((s, i)) else: revisions_details.append((s, i)) else: revisions_details = revisions_ix # Practically, don't compute impacts of revisions prior to first # point that was actually revised if len(revisions_ix) > 0: revisions_details_start = max(revisions_ix[0][0], revisions_details_start) # Setup default (empty) output for revisions revised_endog = None revised_all = None revised_prev_all = None revisions_all = None revised = None revised_prev = None revisions = None revision_weights = None revision_detailed_impacts = None revision_results = None revision_impacts = None # Get revisions datapoints for all revisions (regardless of whether # or not we are computing detailed impacts) if len(revisions_ix) > 0: # Indexes revised_j, revised_p = zip(*revisions_ix) compute_j = np.arange(revised_j[0], revised_j[-1] + 1) # Data from updated model revised_endog = self.endog[:, :previous.nobs].copy() # ("revisions" are points where data was previously published and # then changed, so we need to ignore "updates", which are points # that were not previously published) revised_endog[previous.missing.astype(bool)] = np.nan # subset to revision periods revised_all = revised_endog.T[compute_j] # Data from original model revised_prev_all = previous.endog.T[compute_j] # revision = updated - original revisions_all = (revised_all - revised_prev_all) # Construct a model from which we can create weights for impacts # through `end` # Construct endog for the new model tmp_endog = revised_endog.T.copy() tmp_nobs = max(end, previous.nobs) oos_nobs = tmp_nobs - previous.nobs if oos_nobs > 0: tmp_endog = np.concatenate([ tmp_endog, np.zeros((oos_nobs, self.k_endog)) * np.nan ], axis=0) # Copy time-varying matrices (required by clone) clone_kwargs = {} for key in self.model.shapes.keys(): if key == 'obs': continue mat = getattr(self, key) if mat.shape[-1] > 1: clone_kwargs[key] = mat[..., :tmp_nobs] rev_mod = previous.model.clone(tmp_endog, **clone_kwargs) init = initialization.Initialization.from_results(self) rev_mod.initialize(init) revision_results = rev_mod.smooth() # Get detailed revision weights, impacts, and forecasts if len(revisions_details) > 0: # Indexes for the subset of revisions for which we are # computing detailed impacts compute_j = np.arange(revisions_details_start, revised_j[-1] + 1) # Offset describing revisions for which we are not computing # detailed impacts offset = revisions_details_start - revised_j[0] revised = revised_all[offset:] revised_prev = revised_prev_all[offset:] revisions = revisions_all[offset:] # Compute the weights of the smoothed state vector compute_t = np.arange(start, end) smoothed_state_weights, _, _ = ( tools._compute_smoothed_state_weights( rev_mod, compute_t=compute_t, compute_j=compute_j, compute_prior_weights=False, scale=previous.scale)) # Convert the weights in terms of smoothed forecasts # t, j, m, p, i ZT = rev_mod.design.T if ZT.shape[0] > 1: ZT = ZT[compute_t] # Subset the states used for the impacts if applicable if state_index is not None: ZT = ZT[:, state_index, :] smoothed_state_weights = ( smoothed_state_weights[:, :, state_index]) # Multiplication gives: t, j, m, p * t, j, m, p, k # Sum along axis=2 gives: t, j, p, k # Transpose to: t, j, k, p (i.e. like t, j, m, p but with k # instead of m) revision_weights = np.nansum( smoothed_state_weights[..., None] * ZT[:, None, :, None, :], axis=2).transpose(0, 1, 3, 2) # Multiplication gives: t, j, k, p * t, j, k, p # Sum along axes 1, 3 gives: t, k # This is also a valid way to compute impacts, but it employs # unnecessary multiplications with zeros; it is better to use # the below method that flattens the revision indices before # computing the impacts # revision_detailed_impacts = np.nansum( # revision_weights * revisions[None, :, None, :], # axis=(1, 3)) # Flatten the weights and revisions along the revised j, k # dimensions so that we only retain the actual revision # elements revised_j, revised_p = zip(*[ s for s in revisions_ix if s[0] >= revisions_details_start]) ix_j = revised_j - revised_j[0] # Shape is: t, k, j * p # Note: have to transpose first so that the two advanced # indexes are next to each other, so that "the dimensions from # the advanced indexing operations are inserted into the result # array at the same spot as they were in the initial array" # (see https://numpy.org/doc/stable/user/basics.indexing.html, # "Combining advanced and basic indexing") revision_weights = ( revision_weights.transpose(0, 2, 1, 3)[:, :, ix_j, revised_p]) # Shape is j * k revisions = revisions[ix_j, revised_p] # Shape is t, k revision_detailed_impacts = revision_weights @ revisions # Similarly, flatten the revised and revised_prev series revised = revised[ix_j, revised_p] revised_prev = revised_prev[ix_j, revised_p] # Squeeze if `t` argument used if t is not None: revision_weights = revision_weights[0] revision_detailed_impacts = revision_detailed_impacts[0] # Get total revision impacts revised_impact_forecasts = ( revision_results.smoothed_forecasts[..., start:end]) if end > revision_results.nobs: predict_start = max(start, revision_results.nobs) p = revision_results.predict( start=predict_start, end=end, **extend_kwargs) revised_impact_forecasts = np.concatenate( (revised_impact_forecasts, p.forecasts), axis=1) revision_impacts = (revised_impact_forecasts - prev_impacted_forecasts).T if t is not None: revision_impacts = revision_impacts[0] # Need to also flatten the revisions items that contain all revisions if len(revisions_ix) > 0: revised_j, revised_p = zip(*revisions_ix) ix_j = revised_j - revised_j[0] revisions_all = revisions_all[ix_j, revised_p] revised_all = revised_all[ix_j, revised_p] revised_prev_all = revised_prev_all[ix_j, revised_p] # Now handle updates if len(updates_ix) > 0: # Figure out which time points we need forecast errors for update_t, update_k = zip(*updates_ix) update_start_t = np.min(update_t) update_end_t = np.max(update_t) if revision_results is None: forecasts = previous.predict( start=update_start_t, end=update_end_t + 1, **extend_kwargs).smoothed_forecasts.T else: forecasts = revision_results.predict( start=update_start_t, end=update_end_t + 1).smoothed_forecasts.T realized = self.endog.T[update_start_t:update_end_t + 1] forecasts_error = realized - forecasts # Now subset forecast errors to only the (time, endog) elements # that are updates ix_t = update_t - update_start_t update_realized = realized[ix_t, update_k] update_forecasts = forecasts[ix_t, update_k] update_forecasts_error = forecasts_error[ix_t, update_k] # Get the gains associated with each of the periods if self.design.shape[2] == 1: design = self.design[..., 0][None, ...] elif end <= self.nobs: design = self.design[..., start:end].transpose(2, 0, 1) else: # Note: this case is no longer possible, since above we raise # ValueError for time-varying case with end > self.nobs if design is None: raise ValueError('Model has time-varying design matrix, so' ' an updated time-varying matrix for' ' period `t` is required.') elif design.ndim == 2: design = design[None, ...] else: design = design.transpose(2, 0, 1) state_gain = previous.smoothed_state_gain( updates_ix, start=start, end=end, extend_kwargs=extend_kwargs) # Subset the states used for the impacts if applicable if state_index is not None: design = design[:, :, state_index] state_gain = state_gain[:, state_index] # Compute the gain in terms of observed variables obs_gain = design @ state_gain # Get the news update_impacts = obs_gain @ update_forecasts_error # Squeeze if `t` argument used if t is not None: obs_gain = obs_gain[0] update_impacts = update_impacts[0] else: update_impacts = None update_forecasts = None update_realized = None update_forecasts_error = None obs_gain = None # Results out = SimpleNamespace( # update to forecast of impacted variables from news # = E[y^i | post] - E[y^i | revision] = weight @ news update_impacts=update_impacts, # update to forecast of variables of interest from revisions # = E[y^i | revision] - E[y^i | previous] revision_detailed_impacts=revision_detailed_impacts, # news = A = y^u - E[y^u | previous] news=update_forecasts_error, # revivions y^r(updated) - y^r(previous) for periods in which # detailed impacts were computed revisions=revisions, # revivions y^r(updated) - y^r(previous) revisions_all=revisions_all, # gain matrix = E[y A'] E[A A']^{-1} gain=obs_gain, # weights on observations for the smoothed signal revision_weights=revision_weights, # forecasts of the updated periods used to construct the news # = E[y^u | revised] update_forecasts=update_forecasts, # realizations of the updated periods used to construct the news # = y^u update_realized=update_realized, # revised observations of the periods that were revised and for # which detailed impacts were computed # = y^r_{revised} revised=revised, # revised observations of the periods that were revised # = y^r_{revised} revised_all=revised_all, # previous observations of the periods that were revised and for # which detailed impacts were computed # = y^r_{previous} revised_prev=revised_prev, # previous observations of the periods that were revised # = y^r_{previous} revised_prev_all=revised_prev_all, # previous forecast of the periods of interest, E[y^i | previous] prev_impacted_forecasts=prev_impacted_forecasts, # post. forecast of the periods of interest, E[y^i | post] post_impacted_forecasts=post_impacted_forecasts, # results object associated with the revision revision_results=revision_results, # total impacts from all revisions (both grouped and detailed) revision_impacts=revision_impacts, # list of (x, y) positions of revisions to endog revisions_ix=revisions_ix, # list of (x, y) positions of revisions to endog for which details # of impacts were computed revisions_details=revisions_details, # list of (x, y) positions of revisions to endog for which impacts # were grouped revisions_grouped=revisions_grouped, # period in which revision details start to be computed revisions_details_start=revisions_details_start, # list of (x, y) positions of updates to endog updates_ix=updates_ix, # index of state variables used to compute impacts state_index=state_index) return out def smoothed_state_gain(self, updates_ix, t=None, start=None, end=None, extend_kwargs=None): r""" Cov(\tilde \alpha_{t}, I) Var(I, I)^{-1} where I is a vector of forecast errors associated with `update_indices`. Parameters ---------- updates_ix : list List of indices `(t, i)`, where `t` denotes a zero-indexed time location and `i` denotes a zero-indexed endog variable. """ # Handle `t` if t is not None and (start is not None or end is not None): raise ValueError('Cannot specify both `t` and `start` or `end`.') if t is not None: start = t end = t + 1 # Defaults if start is None: start = self.nobs - 1 if end is None: end = self.nobs if extend_kwargs is None: extend_kwargs = {} # Sanity checks if start < 0 or end < 0: raise ValueError('Negative `t`, `start`, or `end` is not allowed.') if end <= start: raise ValueError('`end` must be after `start`') # Dimensions n_periods = end - start n_updates = len(updates_ix) # Helper to get possibly matrix that is possibly time-varying def get_mat(which, t): mat = getattr(self, which) if mat.shape[-1] > 1: if t < self.nobs: out = mat[..., t] else: if (which not in extend_kwargs or extend_kwargs[which].shape[-1] <= t - self.nobs): raise ValueError(f'Model has time-varying {which}' ' matrix, so an updated time-varying' ' matrix for the extension period is' ' required.') out = extend_kwargs[which][..., t - self.nobs] else: out = mat[..., 0] return out # Helper to get Cov(\tilde \alpha_{t}, I) def get_cov_state_revision(t): tmp1 = np.zeros((self.k_states, n_updates)) for i in range(n_updates): t_i, k_i = updates_ix[i] acov = self.smoothed_state_autocovariance( lag=t - t_i, t=t, extend_kwargs=extend_kwargs) Z_i = get_mat('design', t_i) tmp1[:, i:i + 1] = acov @ Z_i[k_i:k_i + 1].T return tmp1 # Compute Cov(\tilde \alpha_{t}, I) tmp1 = np.zeros((n_periods, self.k_states, n_updates)) for s in range(start, end): tmp1[s - start] = get_cov_state_revision(s) # Compute Var(I) tmp2 = np.zeros((n_updates, n_updates)) for i in range(n_updates): t_i, k_i = updates_ix[i] for j in range(i + 1): t_j, k_j = updates_ix[j] Z_i = get_mat('design', t_i) Z_j = get_mat('design', t_j) acov = self.smoothed_state_autocovariance( lag=t_i - t_j, t=t_i, extend_kwargs=extend_kwargs) tmp2[i, j] = tmp2[j, i] = np.squeeze( Z_i[k_i:k_i + 1] @ acov @ Z_j[k_j:k_j + 1].T ) if t_i == t_j: H = get_mat('obs_cov', t_i) if i == j: tmp2[i, j] += H[k_i, k_j] else: tmp2[i, j] += H[k_i, k_j] tmp2[j, i] += H[k_i, k_j] # Gain gain = tmp1 @ np.linalg.inv(tmp2) if t is not None: gain = gain[0] return gain def _get_smoothed_forecasts(self): if self._smoothed_forecasts is None: # Initialize empty arrays self._smoothed_forecasts = np.zeros(self.forecasts.shape, dtype=self.dtype) self._smoothed_forecasts_error = ( np.zeros(self.forecasts_error.shape, dtype=self.dtype) ) self._smoothed_forecasts_error_cov = ( np.zeros(self.forecasts_error_cov.shape, dtype=self.dtype) ) for t in range(self.nobs): design_t = 0 if self.design.shape[2] == 1 else t obs_cov_t = 0 if self.obs_cov.shape[2] == 1 else t obs_intercept_t = 0 if self.obs_intercept.shape[1] == 1 else t mask = ~self.missing[:, t].astype(bool) # We can recover forecasts self._smoothed_forecasts[:, t] = np.dot( self.design[:, :, design_t], self.smoothed_state[:, t] ) + self.obs_intercept[:, obs_intercept_t] if self.nmissing[t] > 0: self._smoothed_forecasts_error[:, t] = np.nan self._smoothed_forecasts_error[mask, t] = ( self.endog[mask, t] - self._smoothed_forecasts[mask, t] ) self._smoothed_forecasts_error_cov[:, :, t] = np.dot( np.dot(self.design[:, :, design_t], self.smoothed_state_cov[:, :, t]), self.design[:, :, design_t].T ) + self.obs_cov[:, :, obs_cov_t] return ( self._smoothed_forecasts, self._smoothed_forecasts_error, self._smoothed_forecasts_error_cov ) @property def smoothed_forecasts(self): return self._get_smoothed_forecasts()[0] @property def smoothed_forecasts_error(self): return self._get_smoothed_forecasts()[1] @property def smoothed_forecasts_error_cov(self): return self._get_smoothed_forecasts()[2] def get_smoothed_decomposition(self, decomposition_of='smoothed_state', state_index=None): r""" Decompose smoothed output into contributions from observations Parameters ---------- decomposition_of : {"smoothed_state", "smoothed_signal"} The object to perform a decomposition of. If it is set to "smoothed_state", then the elements of the smoothed state vector are decomposed into the contributions of each observation. If it is set to "smoothed_signal", then the predictions of the observation vector based on the smoothed state vector are decomposed. Default is "smoothed_state". state_index : array_like, optional An optional index specifying a subset of states to use when constructing the decomposition of the "smoothed_signal". For example, if `state_index=[0, 1]` is passed, then only the contributions of observed variables to the smoothed signal arising from the first two states will be returned. Note that if not all states are used, the contributions will not sum to the smoothed signal. Default is to use all states. Returns ------- data_contributions : array Contributions of observations to the decomposed object. If the smoothed state is being decomposed, then `data_contributions` are shaped `(nobs, k_states, nobs, k_endog)`, where the `(t, m, j, p)`-th element is the contribution of the `p`-th observation at time `j` to the `m`-th state at time `t`. If the smoothed signal is being decomposed, then `data_contributions` are shaped `(nobs, k_endog, nobs, k_endog)`, where the `(t, k, j, p)`-th element is the contribution of the `p`-th observation at time `j` to the smoothed prediction of the `k`-th observation at time `t`. obs_intercept_contributions : array Contributions of the observation intercept to the decomposed object. If the smoothed state is being decomposed, then `obs_intercept_contributions` are shaped `(nobs, k_states, nobs, k_endog)`, where the `(t, m, j, p)`-th element is the contribution of the `p`-th observation intercept at time `j` to the `m`-th state at time `t`. If the smoothed signal is being decomposed, then `obs_intercept_contributions` are shaped `(nobs, k_endog, nobs, k_endog)`, where the `(t, k, j, p)`-th element is the contribution of the `p`-th observation at time `j` to the smoothed prediction of the `k`-th observation at time `t`. state_intercept_contributions : array Contributions of the state intercept to the decomposed object. If the smoothed state is being decomposed, then `state_intercept_contributions` are shaped `(nobs, k_states, nobs, k_states)`, where the `(t, m, j, l)`-th element is the contribution of the `l`-th state intercept at time `j` to the `m`-th state at time `t`. If the smoothed signal is being decomposed, then `state_intercept_contributions` are shaped `(nobs, k_endog, nobs, k_endog)`, where the `(t, k, j, l)`-th element is the contribution of the `p`-th observation at time `j` to the smoothed prediction of the `k`-th observation at time `t`. prior_contributions : array Contributions of the prior to the decomposed object. If the smoothed state is being decomposed, then `prior_contributions` are shaped `(nobs, k_states, k_states)`, where the `(t, m, l)`-th element is the contribution of the `l`-th element of the prior mean to the `m`-th state at time `t`. If the smoothed signal is being decomposed, then `prior_contributions` are shaped `(nobs, k_endog, k_states)`, where the `(t, k, l)`-th element is the contribution of the `l`-th element of the prior mean to the smoothed prediction of the `k`-th observation at time `t`. Notes ----- Denote the smoothed state at time :math:`t` by :math:`\alpha_t`. Then the smoothed signal is :math:`Z_t \alpha_t`, where :math:`Z_t` is the design matrix operative at time :math:`t`. """ if decomposition_of not in ['smoothed_state', 'smoothed_signal']: raise ValueError('Invalid value for `decomposition_of`. Must be' ' one of "smoothed_state" or "smoothed_signal".') weights, state_intercept_weights, prior_weights = ( tools._compute_smoothed_state_weights( self.model, compute_prior_weights=True, scale=self.scale)) # Get state space objects ZT = self.model.design.T # t, m, p dT = self.model.obs_intercept.T # t, p cT = self.model.state_intercept.T # t, m # Subset the states used for the impacts if applicable if decomposition_of == 'smoothed_signal' and state_index is not None: ZT = ZT[:, state_index, :] weights = weights[:, :, state_index] prior_weights = prior_weights[:, state_index, :] # Convert the weights in terms of smoothed signal # t, j, m, p, i if decomposition_of == 'smoothed_signal': # Multiplication gives: t, j, m, p * t, j, m, p, k # Sum along axis=2 gives: t, j, p, k # Transpose to: t, j, k, p (i.e. like t, j, m, p but with k instead # of m) weights = np.nansum(weights[..., None] * ZT[:, None, :, None, :], axis=2).transpose(0, 1, 3, 2) # Multiplication gives: t, j, m, l * t, j, m, l, k # Sum along axis=2 gives: t, j, l, k # Transpose to: t, j, k, l (i.e. like t, j, m, p but with k instead # of m and l instead of p) state_intercept_weights = np.nansum( state_intercept_weights[..., None] * ZT[:, None, :, None, :], axis=2).transpose(0, 1, 3, 2) # Multiplication gives: t, m, l * t, m, l, k = t, m, l, k # Sum along axis=1 gives: t, l, k # Transpose to: t, k, l (i.e. like t, m, l but with k instead of m) prior_weights = np.nansum( prior_weights[..., None] * ZT[:, :, None, :], axis=1).transpose(0, 2, 1) # Contributions of observations: multiply weights by observations # Multiplication gives t, j, {m,k}, p data_contributions = weights * self.model.endog.T[None, :, None, :] # Transpose to: t, {m,k}, j, p data_contributions = data_contributions.transpose(0, 2, 1, 3) # Contributions of obs intercept: multiply data weights by obs # intercept # Multiplication gives t, j, {m,k}, p obs_intercept_contributions = -weights * dT[None, :, None, :] # Transpose to: t, {m,k}, j, p obs_intercept_contributions = ( obs_intercept_contributions.transpose(0, 2, 1, 3)) # Contributions of state intercept: multiply state intercept weights # by state intercept # Multiplication gives t, j, {m,k}, l state_intercept_contributions = ( state_intercept_weights * cT[None, :, None, :]) # Transpose to: t, {m,k}, j, l state_intercept_contributions = ( state_intercept_contributions.transpose(0, 2, 1, 3)) # Contributions of prior: multiply weights by prior # Multiplication gives t, {m, k}, l prior_contributions = prior_weights * self.initial_state[None, None, :] return (data_contributions, obs_intercept_contributions, state_intercept_contributions, prior_contributions)