import theano.tensor as tt
with pm.Model() as marginal:
    # parameterize the priors with sample mean and standard deviation and a derived metric
    alpha = 1 / data.actual.mean()
    beta = 1 / data.actual.std()
    gamma = 1

    # normal distribution priors - mu and std for before and after
    mu = pm.Exponential('mu', lam=alpha, shape=[2,])
    sd = pm.Exponential('sd', lam=beta, shape=[2,])
    nu = pm.Exponential('nu', lam=gamma, testval=.1)

    # get our time index
    t = np.arange(data.shape[0])

    tau = pm.Beta('tau', alpha=1.1, beta=1.1)
    tau_scaled = tau * t.max()
    switch_proportion = pm.math.sigmoid(tau_scaled - t)
    _mu = (1-switch_proportion) * mu[0] + switch_proportion * mu[1]
    _sd = (1-switch_proportion) * sd[0] + switch_proportion * sd[1]

    # StudentT has other parameters too.
    observation = pm.StudentT('observation', nu=nu, mu=_mu, sd=_sd, observed=data.actual)

with marginal:
    # note large tuning period
    trace_marginal = pm.sample(draws=2000, tune=8000, cores=2)
    pm.traceplot(trace_marginal);