pymc-extras 0.4.1__py3-none-any.whl → 0.6.0__py3-none-any.whl

pymc_extras/statespace/models/DFM.py

@@ -0,0 +1,834 @@
+from collections.abc import Sequence
+from typing import Any
+
+import pytensor
+import pytensor.tensor as pt
+
+from pymc_extras.statespace.core.statespace import PyMCStateSpace
+from pymc_extras.statespace.models.utilities import make_default_coords, validate_names
+from pymc_extras.statespace.utils.constants import (
+    ALL_STATE_AUX_DIM,
+    ALL_STATE_DIM,
+    AR_PARAM_DIM,
+    ERROR_AR_PARAM_DIM,
+    EXOG_STATE_DIM,
+    FACTOR_DIM,
+    OBS_STATE_AUX_DIM,
+    OBS_STATE_DIM,
+    TIME_DIM,
+)
+
+floatX = pytensor.config.floatX
+
+
+class BayesianDynamicFactor(PyMCStateSpace):
+    r"""
+    Dynamic Factor Models
+
+    Notes
+    -----
+    The Dynamic Factor Model (DFM) is a multivariate state-space model used to represent high-dimensional time series
+    as being driven by a smaller set of unobserved dynamic factors.
+
+    Given a set of observed time series :math:`\{y_t\}_{t=0}^T`, where
+
+    .. math::
+        y_t = \begin{bmatrix} y_{1,t} & y_{2,t} & \cdots & y_{k_{\text{endog}},t} \end{bmatrix}^T,
+
+    the DFM assumes that each series is a linear combination of a few latent factors and (optionally) autoregressive errors.
+
+    Let:
+    - :math:`k` be the number of dynamic factors (k_factors),
+    - :math:`p` be the order of the latent factor process (factor_order),
+    - :math:`q` be the order of the observation error process (error_order).
+
+    The model equations are in reduced form is:
+
+    .. math::
+        y_t &= \Lambda f_t + B x_t + u_t + \eta_t \\
+        f_t &= A_1 f_{t-1} + \cdots + A_p f_{t-p} + \varepsilon_{f,t} \\
+        u_t &= C_1 u_{t-1} + \cdots + C_q u_{t-q} + \varepsilon_{u,t}
+
+    Where:
+    - :math:`f_t` is the vector of latent dynamic factors (size :math:`k`),
+    - :math:`x_t` is an optional vector of exogenous variables
+    - :math:`u_t` is a vector of autoregressive observation errors (if `error_var=True` with a VAR(q) structure, else treated as independent AR processes),
+    - :math:`\eta_t \sim \mathcal{N}(0, H_t)` is an optional measurement error (if `measurement_error=True`),
+    - :math:`\varepsilon_{f,t} \sim \mathcal{N}(0, I)` and :math:`\varepsilon_{u,t} \sim \mathcal{N}(0, \Sigma_u)` are independent noise terms.
+        To identify the factors, the innovations to the factor process are standardized with identity covariance.
+
+    Internally, the model is represented in state-space form by stacking all current and lagged latent factors and (if present)
+    AR observation errors into a single state vector of dimension:  :math:: k_{\text{states}} = k \cdot p + k_{\text{endog}} \cdot q,
+    where :math:`k_{\text{endog}}` is the number of observed time series.
+
+    The state vector is defined as:
+
+    .. math::
+        s_t = \begin{bmatrix}
+            f_t(1) \\
+            \vdots \\
+            f_t(k) \\
+            f_{t-p+1}(1) \\
+            \vdots \\
+            f_{t-p+1}(k) \\
+            u_t(1) \\
+            \vdots \\
+            u_t(k_{\text{endog}}) \\
+            \vdots \\
+            u_{t-q+1}(1) \\
+            \vdots \\
+            u_{t-q+1}(k_{\text{endog}})
+        \end{bmatrix}
+        \in \mathbb{R}^{k_{\text{states}}}
+
+    The transition equation is given by:
+
+    .. math::
+        s_{t+1} = T s_t + R \epsilon_t
+
+    Where:
+    - :math:`T` is the state transition matrix, composed of:
+        - VAR coefficients :math:`A_1, \dots, A_{p*k_factors}` for the factors,
+        - (if enabled) AR coefficients :math:`C_1, \dots, C_q` for the observation errors.
+        .. math::
+            T = \begin{bmatrix}
+            A_{1,1}  &   A_{1,2}  &   \cdots  &   A_{1,p}  &   0       &   0       &   \cdots  &   0 \\
+            A_{2,1}  &   A_{2,2}  &   \cdots  &   A_{2,p}  &   0       &   0       &   \cdots  &   0 \\
+            1       &   0       &   \cdots  &   0          &   0       &   0       &   \cdots  &   0 \\
+            0       &   1       &   \cdots  &   0          &   0       &   0       &   \cdots  &   0 \\
+            \vdots  &   \vdots  &   \ddots  &   \vdots     &   \vdots  &   \vdots  &   \ddots  &   \vdots \\
+            \hline
+            0       &   0       &   \cdots  &   0       &   C_{1,1}  &  \cdots  &    C_{1,2} &   C_{1,q} \\
+            0       &   0       &   \cdots  &   0       &   1       &   0       &   \cdots  &   0 \\
+            0       &   0       &   \cdots  &   0       &   0       &   1       &   \cdots  &   0 \\
+            \vdots  &   \vdots  &           &   \vdots  &   \vdots  &   \vdots  &   \ddots  &   \vdots
+            \end{bmatrix}
+            \in \mathbb{R}^{k_{\text{states}} \times k_{\text{states}}}
+
+    - :math:`\epsilon_t` contains the independent shocks (innovations) and has dimension :math:`k + k_{\text{endog}}` if AR errors are included.
+        .. math::
+            \epsilon_t = \begin{bmatrix}
+            \epsilon_{f,t} \\
+            \epsilon_{u,t}
+            \end{bmatrix}
+            \in \mathbb{R}^{k +  k_{\text{endog}}}
+
+    - :math:`R` is a selection matrix mapping shocks to state transitions.
+        .. math::
+            R = \begin{bmatrix}
+            1       &   0       &   \cdots  &   0       &   0       &   0       &   \cdots  &   0 \\
+            0       &   1       &   \cdots  &   0       &   0       &   0       &   \cdots  &   0 \\
+            \vdots  &   \vdots  &   \ddots  &   \vdots  &   \vdots  &   \vdots  &   \ddots  &   \vdots \\
+            0       &   0       &   \cdots  &   1       &   0       &   0       &   \cdots  &   0 \\
+            0       &   0       &   \cdots  &   0       &   1       &   0       &   \cdots  &   0 \\
+            0       &   0       &   \cdots  &   0       &   0       &   1       &   \cdots  &   0 \\
+            \vdots  &   \vdots  &   \ddots  &   \vdots  &   \vdots  &   \vdots  &   \ddots  &   \vdots \\
+            \end{bmatrix}
+            \in \mathbb{R}^{k_{\text{states}} \times (k + k_{\text{endog}})}
+
+    The observation equation is given by:
+
+    .. math::
+
+        y_t = Z s_t + \eta_t
+
+    where
+
+    - :math:`y_t` is the vector of observed variables at time :math:`t`
+
+    - :math:`Z` is the design matrix of the state space representation
+        .. math::
+            Z = \begin{bmatrix}
+            \lambda_{1,1}       &   \lambda_{1,k}   &   \vdots    &   1   &   0   &   \cdots  &   0   &   0   &   \cdots  &   0 \\
+            \lambda_{2,1}       &   \lambda_{2,k}   &   \vdots    &   0   &   1   &   \cdots   &   0   &   \cdots  &   0 \\
+            \vdots              &   \vdots          &   \vdots  &   \vdots  &   \ddots  &   \vdots  &   \vdots  &   \ddots  &   \vdots \\
+            \lambda_{k_{\text{endog}},1}  &   \cdots  &   \lambda_{k_{\text{endog}},k}  &   0   &   0   &   \cdots  &   1   &   0   &   \cdots  &   0 \\
+            \end{bmatrix}
+            \in \mathbb{R}^{k_{\text{endog}} \times k_{\text{states}}}
+
+    - :math:`\eta_t` is the vector of observation errors at time :math:`t`
+
+    When exogenous variables :math:`x_t` are present, the implementation follows `pymc_extras/statespace/models/structural/components/regression.py`.
+    In this case, the state vector is extended to include the beta parameters, and the design matrix is modified accordingly,
+    becoming 3-dimensional to handle time-varying exogenous regressors.
+    This approach provides greater flexibility, controlled by the boolean flags `shared_exog_state` and `exog_innovations`.
+    Unlike Statsmodels, where exogenous variables are included only in the observation equation, here they are fully integrated into the state-space
+    representation.
+
+    .. warning::
+
+        Identification can be an issue, particularly when many observed series load onto only a few latent factors.
+        These models are only identified up to a sign flip in the factor loadings. Proper prior specification is crucial
+        for good estimation and inference.
+
+    Examples
+    --------
+    The following code snippet estimates a dynamic factor model with 1 latent factors,
+    a AR(2) structure on the factor and a AR(1) structure on the errors:
+
+    .. code:: python
+
+        import pymc_extras.statespace as pmss
+        import pymc as pm
+
+        # Create DFM Statespace Model
+        dfm_mod = pmss.BayesianDynamicFactor(
+                k_factors=1,                # Number of latent dynamic factors
+                factor_order=2,             # Number of lags for the latent factor process
+                endog_names=data.columns,   # Names of the observed time series (endogenous variables) (we could also use k_endog = len(data.columns))
+                error_order=1,              # Order of the autoregressive process for the observation noise (i.e., AR(q) error, here q=1)
+                error_var=False,            # If False, models errors as separate AR processes
+                error_cov_type="diagonal",  # Structure of the observation error covariance matrix: uncorrelated noise across series
+                measurement_error=True,     # Whether to include a measurement error term in the model
+                verbose=True
+            )
+
+        # Unpack coords
+        coords = dfm_mod.coords
+
+
+        with pm.Model(coords=coords) as pymc_mod:
+            # Priors for the initial state mean and covariance
+            x0 = pm.Normal("x0", dims=["state_dim"])
+            P0 = pm.HalfNormal("P0", dims=["state_dim", "state_dim"])
+
+            # Factor loadings: shape (k_endog, k_factors)
+            factor_loadings = pm.Normal("factor_loadings", sigma=1, dims=["k_endog", "k_factors"])
+
+            # AR coefficients for factor dynamics: shape (k_factors, factor_order)
+            factor_ar = pm.Normal("factor_ar", sigma=1, dims=["k_factors", "k_factors" * "factor_order"])
+
+            # AR coefficients for observation noise: shape (k_endog, error_order)
+            error_ar = pm.Normal("error_ar", sigma=1, dims=["k_endog", "error_order"])
+
+            # Std devs for observation noise: shape (k_endog,)
+            error_sigma = pm.HalfNormal("error_sigma", dims=["k_endog"])
+
+            # Observation noise covariance matrix
+            obs_sigma = pm.HalfNormal("sigma_obs", dims=["k_endog"])
+
+            # Build the symbolic graph and attach it to the model
+            dfm_mod.build_statespace_graph(data=data, mode="JAX")
+
+            # Sampling
+            idata = pm.sample(
+                draws=500,
+                chains=2,
+                nuts_sampler="nutpie",
+                nuts_sampler_kwargs={"backend": "jax", "gradient_backend": "jax"},
+            )
+
+    """
+
+    def __init__(
+        self,
+        k_factors: int,
+        factor_order: int,
+        endog_names: Sequence[str] | None = None,
+        exog_names: Sequence[str] | None = None,
+        shared_exog_states: bool = False,
+        exog_innovations: bool = False,
+        error_order: int = 0,
+        error_var: bool = False,
+        error_cov_type: str = "diagonal",
+        measurement_error: bool = False,
+        verbose: bool = True,
+    ):
+        """
+        Create a Bayesian Dynamic Factor Model.
+
+        Parameters
+        ----------
+        k_factors : int
+            Number of latent factors.
+
+        factor_order : int
+            Order of the VAR process for the latent factors. If set to 0, the factors have no autoregressive dynamics
+            and are modeled as a white noise process, i.e., :math:`f_t = \varepsilon_{f,t}`.
+            Therefore, the state vector will include one state per factor and "factor_ar" will not exist.
+
+        endog_names : list of str, optional
+            Names of the observed time series.
+
+        exog_names : Sequence[str], optional
+            Names of the exogenous variables.
+
+        shared_exog_states: bool, optional
+            Whether exogenous latent states are shared across the observed states. If True, there will be only one set of exogenous latent
+            states, which are observed by all observed states. If False, each observed state has its own set of exogenous latent states.
+
+        exog_innovations : bool, optional
+            Whether to allow time-varying regression coefficients. If True, coefficients follow a random walk.
+
+        error_order : int, optional
+            Order of the AR process for the observation error component.
+            Default is 0, corresponding to white noise errors.
+
+        error_var : bool, optional
+            If True, errors are modeled jointly via a VAR process;
+            otherwise, each error is modeled separately.
+
+        error_cov_type : {'scalar', 'diagonal', 'unstructured'}, optional
+            Structure of the covariance matrix of the observation errors.
+
+        measurement_error: bool, default True
+            If true, a measurement error term is added to the model.
+
+        verbose: bool, default True
+            If true, a message will be logged to the terminal explaining the variable names, dimensions, and supports.
+
+        """
+
+        validate_names(endog_names, var_name="endog_names", optional=False)
+        k_endog = len(endog_names)
+        self.endog_names = endog_names
+        self.k_endog = k_endog
+        self.k_factors = k_factors
+        self.factor_order = factor_order
+        self.error_order = error_order
+        self.error_var = error_var
+        self.error_cov_type = error_cov_type
+
+        if exog_names is not None:
+            self.shared_exog_states = shared_exog_states
+            self.exog_innovations = exog_innovations
+            validate_names(
+                exog_names, var_name="exog_names", optional=True
+            )  # Not sure if this adds anything
+            k_exog = len(exog_names)
+            self.k_exog = k_exog
+            self.exog_names = exog_names
+        else:
+            self.k_exog = 0
+
+        self.k_exog_states = self.k_exog * self.k_endog if not shared_exog_states else self.k_exog
+        self.exog_flag = self.k_exog > 0
+
+        # Determine the dimension for the latent factor states.
+        # For static factors, one use k_factors.
+        # For dynamic factors with lags, the state include current factors and past lags.
+        # If factor_order is 0, we treat the factor as static (no dynamics),
+        # but it is still included in the state vector with one state per factor. Factor_ar paramter will not exist in this case.
+        k_factor_states = max(self.factor_order, 1) * k_factors
+
+        # Determine the dimension for the error component.
+        # If error_order > 0 then we add additional states for error dynamics, otherwise white noise error.
+        k_error_states = k_endog * error_order if error_order > 0 else 0
+
+        # Total state dimension
+        k_states = k_factor_states + k_error_states + self.k_exog_states
+
+        # Number of independent shocks.
+        # Typically, the latent factors introduce k_factors shocks.
+        # If error_order > 0 and errors are modeled jointly or separately, add appropriate count.
+        k_posdef = k_factors + (k_endog if error_order > 0 else 0) + self.k_exog_states
+        # k_posdef = (k_factors + (k_endog if error_order > 0 else 0) + self.k_exog_states if self.exog_innovations else 0)
+
+        # Initialize the PyMCStateSpace base class.
+        super().__init__(
+            k_endog=k_endog,
+            k_states=k_states,
+            k_posdef=k_posdef,
+            verbose=verbose,
+            measurement_error=measurement_error,
+        )
+
+    @property
+    def param_names(self):
+        names = [
+            "x0",
+            "P0",
+            "factor_loadings",
+            "factor_ar",
+            "error_ar",
+            "error_sigma",
+            "error_cov",
+            "sigma_obs",
+            "beta",
+            "beta_sigma",
+        ]
+
+        # Handle cases where parameters should be excluded based on model settings
+        if self.factor_order == 0:
+            names.remove("factor_ar")
+        if self.error_order == 0:
+            names.remove("error_ar")
+        if self.error_cov_type in ["scalar", "diagonal"]:
+            names.remove("error_cov")
+        if self.error_cov_type == "unstructured":
+            names.remove("error_sigma")
+        if not self.measurement_error:
+            names.remove("sigma_obs")
+        if not self.exog_flag:
+            names.remove("beta")
+            names.remove("beta_sigma")
+        if self.exog_flag and not self.exog_innovations:
+            names.remove("beta_sigma")
+
+        return names
+
+    @property
+    def param_info(self) -> dict[str, dict[str, Any]]:
+        info = {
+            "x0": {
+                "shape": (self.k_states,),
+                "constraints": None,
+            },
+            "P0": {
+                "shape": (self.k_states, self.k_states),
+                "constraints": "Positive Semi-definite",
+            },
+            "factor_loadings": {
+                "shape": (self.k_endog, self.k_factors),
+                "constraints": None,
+            },
+            "factor_ar": {
+                "shape": (self.k_factors, self.factor_order * self.k_factors),
+                "constraints": None,
+            },
+            "error_ar": {
+                "shape": (
+                    self.k_endog,
+                    self.error_order * self.k_endog if self.error_var else self.error_order,
+                ),
+                "constraints": None,
+            },
+            "error_sigma": {
+                "shape": (self.k_endog,) if self.error_cov_type == "diagonal" else (),
+                "constraints": "Positive",
+            },
+            "error_cov": {
+                "shape": (self.k_endog, self.k_endog),
+                "constraints": "Positive Semi-definite",
+            },
+            "sigma_obs": {
+                "shape": (self.k_endog,),
+                "constraints": "Positive",
+            },
+            "beta": {
+                "shape": (self.k_exog_states,),
+                "constraints": None,
+            },
+            "beta_sigma": {
+                "shape": (self.k_exog_states,),
+                "constraints": "Positive",
+            },
+        }
+
+        for name in self.param_names:
+            info[name]["dims"] = self.param_dims[name]
+
+        return {name: info[name] for name in self.param_names}
+
+    @property
+    def state_names(self) -> list[str]:
+        """
+        Returns the names of the hidden states: first factor states (with lags),
+        idiosyncratic error states (with lags), then exogenous states.
+        """
+        names = [
+            f"L{lag}.factor_{i}"
+            for i in range(self.k_factors)
+            for lag in range(max(self.factor_order, 1))
+        ]
+
+        if self.error_order > 0:
+            names.extend(
+                f"L{lag}.error_{i}" for i in range(self.k_endog) for lag in range(self.error_order)
+            )
+
+        if self.exog_flag:
+            if self.shared_exog_states:
+                names.extend([f"beta_{exog_name}[shared]" for exog_name in self.exog_names])
+            else:
+                names.extend(
+                    f"beta_{exog_name}[{endog_name}]"
+                    for exog_name in self.exog_names
+                    for endog_name in self.endog_names
+                )
+        return names
+
+    @property
+    def observed_states(self) -> list[str]:
+        """
+        Returns the names of the observed states (i.e., the endogenous variables).
+        """
+        return self.endog_names
+
+    @property
+    def coords(self) -> dict[str, Sequence]:
+        coords = make_default_coords(self)
+
+        coords[FACTOR_DIM] = [f"factor_{i+1}" for i in range(self.k_factors)]
+
+        if self.factor_order > 0:
+            coords[AR_PARAM_DIM] = list(range(1, (self.factor_order * self.k_factors) + 1))
+
+        if self.error_order > 0:
+            if self.error_var:
+                coords[ERROR_AR_PARAM_DIM] = list(range(1, (self.error_order * self.k_endog) + 1))
+            else:
+                coords[ERROR_AR_PARAM_DIM] = list(range(1, self.error_order + 1))
+
+        if self.exog_flag:
+            coords[EXOG_STATE_DIM] = list(range(1, self.k_exog_states + 1))
+
+        return coords
+
+    @property
+    def shock_names(self) -> list[str]:
+        shock_names = [f"factor_shock_{i}" for i in range(self.k_factors)]
+
+        if self.error_order > 0:
+            shock_names.extend(f"error_shock_{i}" for i in range(self.k_endog))
+
+        if self.exog_flag:
+            if self.shared_exog_states:
+                shock_names.extend(f"exog_shock_{i}.shared" for i in range(self.k_exog))
+            else:
+                shock_names.extend(
+                    f"exog_shock_{i}.endog_{j}"
+                    for i in range(self.k_exog)
+                    for j in range(self.k_endog)
+                )
+
+        return shock_names
+
+    @property
+    def param_dims(self):
+        coord_map = {
+            "x0": (ALL_STATE_DIM,),
+            "P0": (ALL_STATE_DIM, ALL_STATE_AUX_DIM),
+            "factor_loadings": (OBS_STATE_DIM, FACTOR_DIM),
+        }
+        if self.factor_order > 0:
+            coord_map["factor_ar"] = (FACTOR_DIM, AR_PARAM_DIM)
+
+        if self.error_order > 0:
+            coord_map["error_ar"] = (OBS_STATE_DIM, ERROR_AR_PARAM_DIM)
+
+        if self.error_cov_type in ["scalar"]:
+            coord_map["error_sigma"] = ()
+
+        elif self.error_cov_type in ["diagonal"]:
+            coord_map["error_sigma"] = (OBS_STATE_DIM,)
+
+        if self.error_cov_type == "unstructured":
+            coord_map["error_cov"] = (OBS_STATE_DIM, OBS_STATE_AUX_DIM)
+
+        if self.measurement_error:
+            coord_map["sigma_obs"] = (OBS_STATE_DIM,)
+
+        if self.exog_flag:
+            coord_map["beta"] = (EXOG_STATE_DIM,)
+            if self.exog_innovations:
+                coord_map["beta_sigma"] = (EXOG_STATE_DIM,)
+
+        return coord_map
+
+    @property
+    def data_info(self):
+        if self.exog_flag:
+            return {
+                "exog_data": {
+                    "shape": (None, self.k_exog),
+                    "dims": (TIME_DIM, EXOG_STATE_DIM),
+                },
+            }
+        return {}
+
+    @property
+    def data_names(self):
+        if self.exog_flag:
+            return ["exog_data"]
+        return []
+
+    def make_symbolic_graph(self):
+        if not self.exog_flag:
+            x0 = self.make_and_register_variable("x0", shape=(self.k_states,), dtype=floatX)
+        else:
+            initial_factor_loadings = self.make_and_register_variable(
+                "x0", shape=(self.k_states - self.k_exog_states,), dtype=floatX
+            )
+            initial_betas = self.make_and_register_variable(
+                "beta", shape=(self.k_exog_states,), dtype=floatX
+            )
+            x0 = pt.concatenate([initial_factor_loadings, initial_betas], axis=0)
+
+        self.ssm["initial_state", :] = x0
+
+        # Initial covariance
+        P0 = self.make_and_register_variable(
+            "P0", shape=(self.k_states, self.k_states), dtype=floatX
+        )
+        self.ssm["initial_state_cov", :, :] = P0
+
+        # Design matrix (Z)
+        # Construction with block structure:
+        # When factor_order <= 1 and error_order = 0:
+        #   [ A ]               A is the factor loadings matrix with shape (k_endog, k_factors)
+        #
+        # When factor_order > 1, add block of zeros for the factors lags:
+        #   [ A | 0 ]           the zero block has shape (k_endog, k_factors * (factor_order - 1))
+        #
+        # When error_order > 0, add identity matrix and additional zero block for errors lags:
+        #   [ A | 0 | I | 0 ]   I is the identity matrix (k_endog, k_endog) and the final zero block
+        #                       has shape (k_endog, k_endog * (error_order - 1))
+        #
+        # When exog_flag=True, exogenous data (exog_data) is included and the design
+        # matrix becomes 3D with the first dimension indexing time:
+        #   - shared_exog_states=True: exog_data is broadcast across all endogenous series
+        #       → shape (n_timepoints, k_endog, k_exog)
+        #   - shared_exog_states=False: each endogenous series gets its own exog block
+        #       → block-diagonal structure with shape (n_timepoints, k_endog, k_exog * k_endog)
+        # In this case, the base design matrix (factors + errors) is repeated over
+        # time and concatenated with the exogenous block. The final design matrix
+        # has shape (n_timepoints, k_endog, n_columns) and combines all components.
+        factor_loadings = self.make_and_register_variable(
+            "factor_loadings", shape=(self.k_endog, self.k_factors), dtype=floatX
+        )
+        # Add factor loadings (A matrix)
+        matrix_parts = [factor_loadings]
+
+        # Add zero block for the factors lags when factor_order > 1
+        if self.factor_order > 1:
+            matrix_parts.append(
+                pt.zeros((self.k_endog, self.k_factors * (self.factor_order - 1)), dtype=floatX)
+            )
+        # Add identity and zero blocks for error lags when error_order > 0
+        if self.error_order > 0:
+            error_matrix = pt.eye(self.k_endog, dtype=floatX)
+            matrix_parts.append(error_matrix)
+            matrix_parts.append(
+                pt.zeros((self.k_endog, self.k_endog * (self.error_order - 1)), dtype=floatX)
+            )
+        if len(matrix_parts) == 1:
+            design_matrix = factor_loadings * 1.0  # copy to ensure a new PyTensor variable
+            design_matrix.name = "design"
+            # TODO: This is a hack to ensure the design matrix isn't identically equal to the factor_loadings when error_order=0 and factor_order=0
+        else:
+            design_matrix = pt.concatenate(matrix_parts, axis=1)
+            design_matrix.name = "design"
+        # Handle exogenous variables (if any)
+        if self.exog_flag:
+            exog_data = self.make_and_register_data("exog_data", shape=(None, self.k_exog))
+            if self.shared_exog_states:
+                # Shared exogenous states: same exog data is used across all endogenous variables
+                # Shape becomes (n_timepoints, k_endog, k_exog)
+                Z_exog = pt.specify_shape(
+                    pt.join(1, *[pt.expand_dims(exog_data, 1) for _ in range(self.k_endog)]),
+                    (None, self.k_endog, self.k_exog),
+                )
+            else:
+                # Separate exogenous states: each endogenous variable gets its own exog block
+                # Create block-diagonal structure and reshape to (n_timepoints, k_endog, k_exog * k_endog)
+                Z_exog = pt.linalg.block_diag(
+                    *[pt.expand_dims(exog_data, 1) for _ in range(self.k_endog)]
+                )
+                Z_exog = pt.specify_shape(Z_exog, (None, self.k_endog, self.k_exog * self.k_endog))
+
+            # Repeat base design_matrix over time dimension to match exogenous time series
+            n_timepoints = Z_exog.shape[0]
+            design_matrix_time = pt.tile(design_matrix, (n_timepoints, 1, 1))
+            # Concatenate the repeated design matrix with exogenous matrix along the last axis
+            # Final shape: (n_timepoints, k_endog, n_columns + n_exog_columns)
+            design_matrix = pt.concatenate([design_matrix_time, Z_exog], axis=2)
+
+        self.ssm["design"] = design_matrix
+
+        # Transition matrix (T)
+        # Construction with block-diagonal structure:
+        # Each latent component (factors, errors, exogenous states) contributes its own transition block,
+        # and the full transition matrix is assembled with block_diag.
+        #   T = block_diag(A, B, C)
+        #
+        # - Factors (block A):
+        #   If factor_order > 0, the factor AR coefficients are organized into a
+        #   VAR(p) companion matrix of size (k_factors * factor_order, k_factors * factor_order).
+        #   This block shifts lagged factor states and applies AR coefficients.
+        #   If factor_order = 0, a zero matrix is used instead.
+        #
+        # - Errors (block B):
+        #   If error_order > 0:
+        #     * error_var=True → build a full VAR(p) companion matrix (cross-series correlations allowed).
+        #     * error_var=False → build independent AR(p) companion matrices (no cross-series effects).
+        #
+        # - Exogenous states (block C):
+        #   If exog_flag=True, exogenous states are either constant or follow a random walk, modeled with an identity
+        #       transition block of size (k_exog_states, k_exog_states).
+        #
+        # The final transition matrix is block-diagonal, combining all active components:
+        #   Transition = block_diag(Factors, Errors, Exogenous)
+
+        # auxiliary functions to build transition matrix block
+        def build_var_block_matrix(ar_coeffs, k_series, p):
+            """
+            Build the VAR(p) companion matrix for the factors.
+
+            ar_coeffs: PyTensor matrix of shape (k_series, p * k_series)
+                    [A1 | A2 | ... | Ap] horizontally concatenated.
+            k_series: number of series
+            p: lag order
+            """
+            size = k_series * p
+            block = pt.zeros((size, size), dtype=floatX)
+
+            # First block row: the AR coefficient matrices for each lag
+            block = block[0:k_series, 0 : k_series * p].set(ar_coeffs)
+
+            # Sub-diagonal identity blocks (shift structure)
+            if p > 1:
+                # Create the identity pattern for all sub-diagonal blocks
+                identity_pattern = pt.eye(k_series * (p - 1), dtype=floatX)
+                block = block[k_series:, : k_series * (p - 1)].set(identity_pattern)
+
+            return block
+
+        def build_independent_var_block_matrix(ar_coeffs, k_series, p):
+            """
+            Build a VAR(p)-style companion matrix for independent AR(p) processes
+            with interleaved state ordering:
+            (x1(t), x2(t), ..., x1(t-1), x2(t-1), ...).
+
+            ar_coeffs: PyTensor matrix of shape (k_series, p)
+            k_series: number of independent series
+            p: lag order
+            """
+            size = k_series * p
+            block = pt.zeros((size, size), dtype=floatX)
+
+            # First block row: AR coefficients per series (block diagonal)
+            for j in range(k_series):
+                for lag in range(p):
+                    col_idx = lag * k_series + j
+                    block = pt.set_subtensor(block[j, col_idx], ar_coeffs[j, lag])
+
+            # Sub-diagonal identity blocks (shift)
+            if p > 1:
+                identity_pattern = pt.eye(k_series * (p - 1), dtype=floatX)
+                block = pt.set_subtensor(block[k_series:, : k_series * (p - 1)], identity_pattern)
+            return block
+
+        transition_blocks = []
+        # Block A: Factors
+        if self.factor_order > 0:
+            factor_ar = self.make_and_register_variable(
+                "factor_ar",
+                shape=(self.k_factors, self.factor_order * self.k_factors),
+                dtype=floatX,
+            )
+            transition_blocks.append(
+                build_var_block_matrix(factor_ar, self.k_factors, self.factor_order)
+            )
+        else:
+            transition_blocks.append(pt.zeros((self.k_factors, self.k_factors), dtype=floatX))
+        # Block B: Errors
+        if self.error_order > 0 and self.error_var:
+            error_ar = self.make_and_register_variable(
+                "error_ar", shape=(self.k_endog, self.error_order * self.k_endog), dtype=floatX
+            )
+            transition_blocks.append(
+                build_var_block_matrix(error_ar, self.k_endog, self.error_order)
+            )
+        elif self.error_order > 0 and not self.error_var:
+            error_ar = self.make_and_register_variable(
+                "error_ar", shape=(self.k_endog, self.error_order), dtype=floatX
+            )
+            transition_blocks.append(
+                build_independent_var_block_matrix(error_ar, self.k_endog, self.error_order)
+            )
+        # Block C: Exogenous states
+        if self.exog_flag:
+            transition_blocks.append(pt.eye(self.k_exog_states, dtype=floatX))
+
+        self.ssm["transition", :, :] = pt.linalg.block_diag(*transition_blocks)
+
+        # Selection matrix (R)
+        for i in range(self.k_factors):
+            self.ssm["selection", i, i] = 1.0
+
+        if self.error_order > 0:
+            for i in range(self.k_endog):
+                row = max(self.factor_order, 1) * self.k_factors + i
+                col = self.k_factors + i
+                self.ssm["selection", row, col] = 1.0
+
+        if self.exog_flag and self.exog_innovations:
+            row_start = self.k_states - self.k_exog_states
+            row_end = self.k_states
+
+            if self.error_order > 0:
+                col_start = self.k_factors + self.k_endog
+                col_end = self.k_factors + self.k_endog + self.k_exog_states
+            else:
+                col_start = self.k_factors
+                col_end = self.k_factors + self.k_exog_states
+
+            self.ssm["selection", row_start:row_end, col_start:col_end] = pt.eye(
+                self.k_exog_states, dtype=floatX
+            )
+
+        factor_cov = pt.eye(self.k_factors, dtype=floatX)
+
+        # Handle error_sigma and error_cov depending on error_cov_type
+        if self.error_cov_type == "scalar":
+            error_sigma = self.make_and_register_variable("error_sigma", shape=(), dtype=floatX)
+            error_cov = pt.eye(self.k_endog) * error_sigma
+        elif self.error_cov_type == "diagonal":
+            error_sigma = self.make_and_register_variable(
+                "error_sigma", shape=(self.k_endog,), dtype=floatX
+            )
+            error_cov = pt.diag(error_sigma)
+        elif self.error_cov_type == "unstructured":
+            error_cov = self.make_and_register_variable(
+                "error_cov", shape=(self.k_endog, self.k_endog), dtype=floatX
+            )
+
+        # State covariance matrix (Q)
+        if self.error_order > 0:
+            if self.exog_flag and self.exog_innovations:
+                # Include AR noise in state vector
+                beta_sigma = self.make_and_register_variable(
+                    "beta_sigma", shape=(self.k_exog_states,), dtype=floatX
+                )
+                exog_cov = pt.diag(beta_sigma)
+                self.ssm["state_cov", :, :] = pt.linalg.block_diag(factor_cov, error_cov, exog_cov)
+            elif self.exog_flag and not self.exog_innovations:
+                exog_cov = pt.zeros((self.k_exog_states, self.k_exog_states), dtype=floatX)
+                self.ssm["state_cov", :, :] = pt.linalg.block_diag(factor_cov, error_cov, exog_cov)
+            elif not self.exog_flag:
+                self.ssm["state_cov", :, :] = pt.linalg.block_diag(factor_cov, error_cov)
+        else:
+            if self.exog_flag and self.exog_innovations:
+                beta_sigma = self.make_and_register_variable(
+                    "beta_sigma", shape=(self.k_exog_states,), dtype=floatX
+                )
+                exog_cov = pt.diag(beta_sigma)
+                self.ssm["state_cov", :, :] = pt.linalg.block_diag(factor_cov, exog_cov)
+            elif self.exog_flag and not self.exog_innovations:
+                exog_cov = pt.zeros((self.k_exog_states, self.k_exog_states), dtype=floatX)
+                self.ssm["state_cov", :, :] = pt.linalg.block_diag(factor_cov, exog_cov)
+            elif not self.exog_flag:
+                # Only latent factor in the state
+                self.ssm["state_cov", :, :] = factor_cov
+
+        # Observation covariance matrix (H)
+        if self.error_order > 0:
+            if self.measurement_error:
+                sigma_obs = self.make_and_register_variable(
+                    "sigma_obs", shape=(self.k_endog,), dtype=floatX
+                )
+                self.ssm["obs_cov", :, :] = pt.diag(sigma_obs)
+            # else: obs_cov remains zero (no measurement noise and idiosyncratic noise captured in state)
+        else:
+            if self.measurement_error:
+                # TODO: check this decision
+                # in this case error_order = 0, so there is no error term in the state, so the sigma error could not be added there
+                # Idiosyncratic + measurement error
+                sigma_obs = self.make_and_register_variable(
+                    "sigma_obs", shape=(self.k_endog,), dtype=floatX
+                )
+                total_obs_var = error_sigma**2 + sigma_obs**2
+                self.ssm["obs_cov", :, :] = pt.diag(pt.sqrt(total_obs_var))
+            else:
+                self.ssm["obs_cov", :, :] = pt.diag(error_sigma)