pymc-extras 0.3.1__py3-none-any.whl → 0.4.0__py3-none-any.whl

pymc_extras/statespace/models/structural/components/seasonality.py

@@ -0,0 +1,445 @@
+import numpy as np
+
+from pytensor import tensor as pt
+
+from pymc_extras.statespace.models.structural.core import Component
+from pymc_extras.statespace.models.structural.utils import _frequency_transition_block
+
+
+class TimeSeasonality(Component):
+    r"""
+    Seasonal component, modeled in the time domain
+
+    Parameters
+    ----------
+    season_length: int
+        The number of periods in a single seasonal cycle, e.g. 12 for monthly data with annual seasonal pattern, 7 for
+        daily data with weekly seasonal pattern, etc.
+
+    innovations: bool, default True
+        Whether to include stochastic innovations in the strength of the seasonal effect
+
+    name: str, default None
+        A name for this seasonal component. Used to label dimensions and coordinates. Useful when multiple seasonal
+        components are included in the same model. Default is ``f"Seasonal[s={season_length}]"``
+
+    state_names: list of str, default None
+        List of strings for seasonal effect labels. If provided, it must be of length ``season_length``. An example
+        would be ``state_names = ['Mon', 'Tue', 'Wed', 'Thur', 'Fri', 'Sat', 'Sun']`` when data is daily with a weekly
+        seasonal pattern (``season_length = 7``).
+
+        If None, states will be numbered ``[State_0, ..., State_s]``
+
+    remove_first_state: bool, default True
+        If True, the first state will be removed from the model. This is done because there are only n-1 degrees of
+        freedom in the seasonal component, and one state is not identified. If False, the first state will be
+        included in the model, but it will not be identified -- you will need to handle this in the priors (e.g. with
+        ZeroSumNormal).
+
+    observed_state_names: list[str] | None, default None
+        List of strings for observed state labels. If None, defaults to ["data"].
+
+    Notes
+    -----
+    A seasonal effect is any pattern that repeats every fixed interval. Although there are many possible ways to
+    model seasonal effects, the implementation used here is the one described by [1] as the "canonical" time domain
+    representation. The seasonal component can be expressed:
+
+    .. math::
+        \gamma_t = -\sum_{i=1}^{s-1} \gamma_{t-i} + \omega_t, \quad \omega_t \sim N(0, \sigma_\gamma)
+
+    Where :math:`s` is the ``seasonal_length`` parameter and :math:`\omega_t` is the (optional) stochastic innovation.
+    To give interpretation to the :math:`\gamma` terms, it is helpful to work  through the algebra for a simple
+    example. Let :math:`s=4`, and omit the shock term. Define initial conditions :math:`\gamma_0, \gamma_{-1},
+    \gamma_{-2}`. The value of the seasonal component for the first 5 timesteps will be:
+
+    .. math::
+        \begin{align}
+            \gamma_1 &= -\gamma_0 - \gamma_{-1} - \gamma_{-2} \\
+             \gamma_2 &= -\gamma_1 - \gamma_0 - \gamma_{-1} \\
+                       &= -(-\gamma_0 - \gamma_{-1} - \gamma_{-2}) - \gamma_0 - \gamma_{-1}  \\
+                       &= (\gamma_0 - \gamma_0 )+ (\gamma_{-1} - \gamma_{-1}) + \gamma_{-2} \\
+                       &= \gamma_{-2} \\
+              \gamma_3 &= -\gamma_2 - \gamma_1 - \gamma_0  \\
+                       &= -\gamma_{-2} - (-\gamma_0 - \gamma_{-1} - \gamma_{-2}) - \gamma_0 \\
+                       &=  (\gamma_{-2} - \gamma_{-2}) + \gamma_{-1} + (\gamma_0 - \gamma_0) \\
+                       &= \gamma_{-1} \\
+              \gamma_4 &= -\gamma_3 - \gamma_2 - \gamma_1 \\
+                       &= -\gamma_{-1} - \gamma_{-2} -(-\gamma_0 - \gamma_{-1} - \gamma_{-2}) \\
+                       &= (\gamma_{-2} - \gamma_{-2}) + (\gamma_{-1} - \gamma_{-1}) + \gamma_0 \\
+                       &= \gamma_0 \\
+              \gamma_5 &= -\gamma_4 - \gamma_3 - \gamma_2 \\
+                       &= -\gamma_0 - \gamma_{-1} - \gamma_{-2} \\
+                       &= \gamma_1
+        \end{align}
+
+    This exercise shows that, given a list ``initial_conditions`` of length ``s-1``, the effects of this model will be:
+
+        - Period 1: ``-sum(initial_conditions)``
+        - Period 2: ``initial_conditions[-1]``
+        - Period 3: ``initial_conditions[-2]``
+        - ...
+        - Period s: ``initial_conditions[0]``
+        - Period s+1: ``-sum(initial_condition)``
+
+    And so on. So for interpretation, the ``season_length - 1`` initial states are, when reversed, the coefficients
+    associated with ``state_names[1:]``.
+
+    .. warning::
+        Although the ``state_names`` argument expects a list of length ``season_length``, only ``state_names[1:]``
+        will be saved as model dimensions, since the 1st coefficient is not identified (it is defined as
+        :math:`-\sum_{i=1}^{s} \gamma_{t-i}`).
+
+    Examples
+    --------
+    Estimate monthly with a model with a gaussian random walk trend and monthly seasonality:
+
+    .. code:: python
+
+        from pymc_extras.statespace import structural as st
+        import pymc as pm
+        import pytensor.tensor as pt
+        import pandas as pd
+
+        # Get month names
+        state_names = pd.date_range('1900-01-01', '1900-12-31', freq='MS').month_name().tolist()
+
+        # Build the structural model
+        grw = st.LevelTrendComponent(order=1, innovations_order=1)
+        annual_season = st.TimeSeasonality(
+            season_length=12, name="annual", state_names=state_names, innovations=False
+        )
+        ss_mod = (grw + annual_season).build()
+
+        with pm.Model(coords=ss_mod.coords) as model:
+            P0 = pm.Deterministic('P0', pt.eye(ss_mod.k_states) * 10, dims=ss_mod.param_dims['P0'])
+
+            initial_level_trend = pm.Deterministic(
+                "initial_level_trend", pt.zeros(1), dims=ss_mod.param_dims["initial_level_trend"]
+            )
+            sigma_level_trend = pm.HalfNormal(
+                "sigma_level_trend", sigma=1e-6, dims=ss_mod.param_dims["sigma_level_trend"]
+            )
+            coefs_annual = pm.Normal("coefs_annual", sigma=1e-2, dims=ss_mod.param_dims["coefs_annual"])
+
+            ss_mod.build_statespace_graph(data)
+            idata = pm.sample(
+                nuts_sampler="nutpie", nuts_sampler_kwargs={"backend": "JAX", "gradient_backend": "JAX"}
+            )
+
+    References
+    ----------
+    .. [1] Durbin, James, and Siem Jan Koopman. 2012.
+        Time Series Analysis by State Space Methods: Second Edition.
+        Oxford University Press.
+    """
+
+    def __init__(
+        self,
+        season_length: int,
+        innovations: bool = True,
+        name: str | None = None,
+        state_names: list | None = None,
+        remove_first_state: bool = True,
+        observed_state_names: list[str] | None = None,
+    ):
+        if observed_state_names is None:
+            observed_state_names = ["data"]
+
+        if name is None:
+            name = f"Seasonal[s={season_length}]"
+        if state_names is None:
+            state_names = [f"{name}_{i}" for i in range(season_length)]
+        else:
+            if len(state_names) != season_length:
+                raise ValueError(
+                    f"state_names must be a list of length season_length, got {len(state_names)}"
+                )
+            state_names = state_names.copy()
+
+        self.innovations = innovations
+        self.remove_first_state = remove_first_state
+
+        if self.remove_first_state:
+            # In traditional models, the first state isn't identified, so we can help out the user by automatically
+            # discarding it.
+            # TODO: Can this be stashed and reconstructed automatically somehow?
+            state_names.pop(0)
+
+        self.provided_state_names = state_names
+
+        k_states = season_length - int(self.remove_first_state)
+        k_endog = len(observed_state_names)
+        k_posdef = int(innovations)
+
+        super().__init__(
+            name=name,
+            k_endog=k_endog,
+            k_states=k_states * k_endog,
+            k_posdef=k_posdef * k_endog,
+            observed_state_names=observed_state_names,
+            measurement_error=False,
+            combine_hidden_states=True,
+            obs_state_idxs=np.tile(np.array([1.0] + [0.0] * (k_states - 1)), k_endog),
+        )
+
+    def populate_component_properties(self):
+        k_states = self.k_states // self.k_endog
+        k_endog = self.k_endog
+
+        self.state_names = [
+            f"{state_name}[{endog_name}]"
+            for endog_name in self.observed_state_names
+            for state_name in self.provided_state_names
+        ]
+        self.param_names = [f"coefs_{self.name}"]
+
+        self.param_info = {
+            f"coefs_{self.name}": {
+                "shape": (k_states,) if k_endog == 1 else (k_endog, k_states),
+                "constraints": None,
+                "dims": (f"state_{self.name}",)
+                if k_endog == 1
+                else (f"endog_{self.name}", f"state_{self.name}"),
+            }
+        }
+
+        self.param_dims = {
+            f"coefs_{self.name}": (f"state_{self.name}",)
+            if k_endog == 1
+            else (f"endog_{self.name}", f"state_{self.name}")
+        }
+
+        self.coords = (
+            {f"state_{self.name}": self.provided_state_names}
+            if k_endog == 1
+            else {
+                f"endog_{self.name}": self.observed_state_names,
+                f"state_{self.name}": self.provided_state_names,
+            }
+        )
+
+        if self.innovations:
+            self.param_names += [f"sigma_{self.name}"]
+            self.param_info[f"sigma_{self.name}"] = {
+                "shape": (),
+                "constraints": "Positive",
+                "dims": None,
+            }
+            self.shock_names = [f"{self.name}[{name}]" for name in self.observed_state_names]
+
+    def make_symbolic_graph(self) -> None:
+        k_states = self.k_states // self.k_endog
+        k_posdef = self.k_posdef // self.k_endog
+        k_endog = self.k_endog
+
+        if self.remove_first_state:
+            # In this case, parameters are normalized to sum to zero, so the current state is the negative sum of
+            # all previous states.
+            T = np.eye(k_states, k=-1)
+            T[0, :] = -1
+        else:
+            # In this case we assume the user to be responsible for ensuring the states sum to zero, so T is just a
+            # circulant matrix that cycles between the states.
+            T = np.eye(k_states, k=1)
+            T[-1, 0] = 1
+
+        self.ssm["transition", :, :] = pt.linalg.block_diag(*[T for _ in range(k_endog)])
+
+        Z = pt.zeros((1, k_states))[0, 0].set(1)
+        self.ssm["design", :, :] = pt.linalg.block_diag(*[Z for _ in range(k_endog)])
+
+        initial_states = self.make_and_register_variable(
+            f"coefs_{self.name}", shape=(k_states,) if k_endog == 1 else (k_endog, k_states)
+        )
+        self.ssm["initial_state", :] = initial_states.ravel()
+
+        if self.innovations:
+            R = pt.zeros((k_states, k_posdef))[0, 0].set(1.0)
+            self.ssm["selection", :, :] = pt.join(0, *[R for _ in range(k_endog)])
+            season_sigma = self.make_and_register_variable(
+                f"sigma_{self.name}", shape=() if k_endog == 1 else (k_endog,)
+            )
+            cov_idx = ("state_cov", *np.diag_indices(k_posdef * k_endog))
+            self.ssm[cov_idx] = season_sigma**2
+
+
+class FrequencySeasonality(Component):
+    r"""
+    Seasonal component, modeled in the frequency domain
+
+    Parameters
+    ----------
+    season_length: float
+        The number of periods in a single seasonal cycle, e.g. 12 for monthly data with annual seasonal pattern, 7 for
+        daily data with weekly seasonal pattern, etc. Non-integer seasonal_length is also permitted, for example
+        365.2422 days in a (solar) year.
+
+    n: int
+        Number of fourier features to include in the seasonal component. Default is ``season_length // 2``, which
+        is the maximum possible. A smaller number can be used for a more wave-like seasonal pattern.
+
+    name: str, default None
+        A name for this seasonal component. Used to label dimensions and coordinates. Useful when multiple seasonal
+        components are included in the same model. Default is ``f"Seasonal[s={season_length}, n={n}]"``
+
+    innovations: bool, default True
+        Whether to include stochastic innovations in the strength of the seasonal effect
+
+    observed_state_names: list[str] | None, default None
+        List of strings for observed state labels. If None, defaults to ["data"].
+
+    Notes
+    -----
+    A seasonal effect is any pattern that repeats every fixed interval. Although there are many possible ways to
+    model seasonal effects, the implementation used here is the one described by [1] as the "canonical" frequency domain
+    representation. The seasonal component can be expressed:
+
+    .. math::
+        \begin{align}
+            \gamma_t &= \sum_{j=1}^{2n} \gamma_{j,t} \\
+            \gamma_{j, t+1} &= \gamma_{j,t} \cos \lambda_j + \gamma_{j,t}^\star \sin \lambda_j + \omega_{j, t} \\
+            \gamma_{j, t}^\star &= -\gamma_{j,t} \sin \lambda_j + \gamma_{j,t}^\star \cos \lambda_j + \omega_{j,t}^\star
+            \lambda_j &= \frac{2\pi j}{s}
+        \end{align}
+
+    Where :math:`s` is the ``seasonal_length``.
+
+    Unlike a ``TimeSeasonality`` component, a ``FrequencySeasonality`` component does not require integer season
+    length. In addition, for long seasonal periods, it is possible to obtain a more compact state space representation
+    by choosing ``n << s // 2``. Using ``TimeSeasonality``, an annual seasonal pattern in daily data requires 364
+    states, whereas ``FrequencySeasonality`` always requires ``2 * n`` states, regardless of the ``seasonal_length``.
+    The price of this compactness is less representational power. At ``n = 1``, the seasonal pattern will be a pure
+    sine wave. At ``n = s // 2``, any arbitrary pattern can be represented.
+
+    One cost of the added flexibility of ``FrequencySeasonality`` is reduced interpretability. States of this model are
+    coefficients :math:`\gamma_1, \gamma^\star_1, \gamma_2, \gamma_2^\star ..., \gamma_n, \gamma^\star_n` associated
+    with different frequencies in the fourier representation of the seasonal pattern. As a result, it is not possible
+    to isolate and identify a "Monday" effect, for instance.
+    """
+
+    def __init__(
+        self,
+        season_length,
+        n=None,
+        name=None,
+        innovations=True,
+        observed_state_names: list[str] | None = None,
+    ):
+        if observed_state_names is None:
+            observed_state_names = ["data"]
+
+        k_endog = len(observed_state_names)
+
+        if n is None:
+            n = int(season_length / 2)
+        if name is None:
+            name = f"Frequency[s={season_length}, n={n}]"
+
+        k_states = n * 2
+        self.n = n
+        self.season_length = season_length
+        self.innovations = innovations
+
+        # If the model is completely saturated (n = s // 2), the last state will not be identified, so it shouldn't
+        # get a parameter assigned to it and should just be fixed to zero.
+        # Test this way (rather than n == s // 2) to catch cases when n is non-integer.
+        self.last_state_not_identified = self.season_length / self.n == 2.0
+        self.n_coefs = k_states - int(self.last_state_not_identified)
+
+        obs_state_idx = np.zeros(k_states)
+        obs_state_idx[slice(0, k_states, 2)] = 1
+        obs_state_idx = np.tile(obs_state_idx, k_endog)
+
+        super().__init__(
+            name=name,
+            k_endog=k_endog,
+            k_states=k_states * k_endog,
+            k_posdef=k_states * int(self.innovations) * k_endog,
+            observed_state_names=observed_state_names,
+            measurement_error=False,
+            combine_hidden_states=True,
+            obs_state_idxs=obs_state_idx,
+        )
+
+    def make_symbolic_graph(self) -> None:
+        k_endog = self.k_endog
+        k_states = self.k_states // k_endog
+        k_posdef = self.k_posdef // k_endog
+        n_coefs = self.n_coefs
+
+        Z = pt.zeros((1, k_states))[0, slice(0, k_states, 2)].set(1.0)
+
+        self.ssm["design", :, :] = pt.linalg.block_diag(*[Z for _ in range(k_endog)])
+
+        init_state = self.make_and_register_variable(
+            f"{self.name}", shape=(n_coefs,) if k_endog == 1 else (k_endog, n_coefs)
+        )
+
+        init_state_idx = np.concatenate(
+            [
+                np.arange(k_states * i, (i + 1) * k_states, dtype=int)[:n_coefs]
+                for i in range(k_endog)
+            ],
+            axis=0,
+        )
+
+        self.ssm["initial_state", init_state_idx] = init_state.ravel()
+
+        T_mats = [_frequency_transition_block(self.season_length, j + 1) for j in range(self.n)]
+        T = pt.linalg.block_diag(*T_mats)
+        self.ssm["transition", :, :] = pt.linalg.block_diag(*[T for _ in range(k_endog)])
+
+        if self.innovations:
+            sigma_season = self.make_and_register_variable(
+                f"sigma_{self.name}", shape=() if k_endog == 1 else (k_endog,)
+            )
+            self.ssm["selection", :, :] = pt.eye(self.k_states)
+            self.ssm["state_cov", :, :] = pt.eye(self.k_posdef) * pt.repeat(
+                sigma_season**2, k_posdef
+            )
+
+    def populate_component_properties(self):
+        k_endog = self.k_endog
+        n_coefs = self.n_coefs
+        k_states = self.k_states // k_endog
+
+        self.state_names = [
+            f"{f}_{self.name}_{i}[{obs_state_name}]"
+            for obs_state_name in self.observed_state_names
+            for i in range(self.n)
+            for f in ["Cos", "Sin"]
+        ]
+        self.param_names = [f"{self.name}"]
+
+        self.param_dims = {self.name: (f"state_{self.name}",)}
+        self.param_info = {
+            f"{self.name}": {
+                "shape": (n_coefs,) if k_endog == 1 else (k_endog, n_coefs),
+                "constraints": None,
+                "dims": (f"state_{self.name}",)
+                if k_endog == 1
+                else (f"endog_{self.name}", f"state_{self.name}"),
+            }
+        }
+
+        # Regardless of whether the fourier basis are saturated, there will always be one symbolic state per basis.
+        # That's why the self.states is just a simple loop over everything. But when saturated, one of those states
+        # doesn't have an associated **parameter**, so the coords need to be adjusted to reflect this.
+        init_state_idx = np.concatenate(
+            [
+                np.arange(k_states * i, (i + 1) * k_states, dtype=int)[:n_coefs]
+                for i in range(k_endog)
+            ],
+            axis=0,
+        )
+        self.coords = {f"state_{self.name}": [self.state_names[i] for i in init_state_idx]}
+
+        if self.innovations:
+            self.shock_names = self.state_names.copy()
+            self.param_names += [f"sigma_{self.name}"]
+            self.param_info[f"sigma_{self.name}"] = {
+                "shape": () if k_endog == 1 else (k_endog, n_coefs),
+                "constraints": "Positive",
+                "dims": None if k_endog == 1 else (f"endog_{self.name}",),
+            }