BUG: SARIMAX with time-varying coefficients works when model is (1,1,0) or (0,1,1), but raises an error when model set to (0,1,0)

[fcela]

Call used:

SARIMAX(endog=endog, exog=exog, order=order
    seasonal_order= (0,0,0,4),
    trend= 'n',
enforce_stationarity=True,
enforce_invertibility=True,
time_varying_regression=True,
concentrate_scale = False,
measurement_error = False,
mle_regression = False)

Endog has 6 covariates.

Works when order = (1,1,0) (however state.1 is all zero); also works for order = (0,1,1). However when using order=(1,1,0) I get the following error:

    524     self.ssm._time_invariant = False
    526 # Initialize the fixed components of the statespace model
--> 527 self.ssm['design'] = self.initial_design
    528 self.ssm['state_intercept'] = self.initial_state_intercept
    529 self.ssm['transition'] = self.initial_transition

File C:\Developer\venv\lab\lib\site-packages\statsmodels\tsa\statespace\representation.py:420, in Representation.__setitem__(self, key, value)
    417     if key not in self.shapes:
    418         raise IndexError('"%s" is an invalid state space matrix name'
    419                          % key)
--> 420     setattr(self, key, value)
    421 # If it's a tuple (with a string as the first element) then we must be
    422 # setting a slice of a matrix
    423 elif _type is tuple:

File C:\Developer\venv\lab\lib\site-packages\statsmodels\tsa\statespace\representation.py:54, in MatrixWrapper.__set__(self, obj, value)
     51 shape = obj.shapes[self.attribute]
     53 if len(shape) == 3:
---> 54     value = self._set_matrix(obj, value, shape)
     55 else:
     56     value = self._set_vector(obj, value, shape)

File C:\Developer\venv\lab\lib\site-packages\statsmodels\tsa\statespace\representation.py:68, in MatrixWrapper._set_matrix(self, obj, value, shape)
     65     value = value[None, :]
     67 # Enforce that the matrix is appropriate size
---> 68 validate_matrix_shape(
     69     self.name, value.shape, shape[0], shape[1], obj.nobs
     70 )
     72 # Expand time-invariant matrix
     73 if value.ndim == 2:

File C:\Developer\venv\lab\lib\site-packages\statsmodels\tsa\statespace\tools.py:1477, in validate_matrix_shape(name, shape, nrows, ncols, nobs)
   1474     raise ValueError('Invalid dimensions for %s matrix: requires %d'
   1475                      ' rows, got %d' % (name, nrows, shape[0]))
   1476 if not shape[1] == ncols:
-> 1477     raise ValueError('Invalid dimensions for %s matrix: requires %d'
   1478                      ' columns, got %d' % (name, ncols, shape[1]))
   1480 # If we do not yet know `nobs`, do not allow time-varying arrays
   1481 if nobs is None and not (ndim == 2 or shape[-1] == 1):

ValueError: Invalid dimensions for design matrix: requires 7 columns, got 6

[ChadFulton]

Thanks for reporting this!

[ChadFulton]

Here's a minimum reproducible example:

import numpy as np
import statsmodels.api as sm

np.random.seed(1234)
endog = np.random.normal(size=100)
exog = np.random.normal(size=(100, 1))

mod = sm.tsa.SARIMAX(endog=endog, exog=exog, order=(0, 1, 0), trend='n',
                     time_varying_regression=True, mle_regression=False)

res = mod.fit(disp=False)