def testMarkovChainLogprobMatchesOriginal(self):
# Use a very short dataset since the non-markov-chain model is slow.
dataset = data.sp500_returns(num_points=10)
model = stochastic_volatility.StochasticVolatility(
use_markov_chain=False, **dataset)
markov_chain_model = stochastic_volatility.StochasticVolatility(
use_markov_chain=True, **dataset)
xs = self.evaluate(model.prior_distribution().sample(
10, seed=tfp_test_util.test_seed()))
xs_stacked = xs[:-1] + (tf.stack(xs[-1], axis=-1), )
self.assertAllClose(
model.unnormalized_log_prob(xs),
markov_chain_model.unnormalized_log_prob(xs_stacked),
atol=1e-2)
def testBasic(self):
"""Checks that unconstrained parameters yield finite joint densities."""
model = stochastic_volatility.StochasticVolatility(
centered_returns=tf.convert_to_tensor([5., -2.1, 8., 4., 1.1]))
self.validate_log_prob_and_transforms(
model,
sample_transformation_shapes=dict(
identity={
'persistence_of_volatility': [],
'mean_log_volatility': [],
'white_noise_shock_scale': [],
'log_volatility': [5]
}))