pymc3 needs a likelihood function to sample, and if you observed a function of some random variables (ie., observing a Deterministic) the likelihood expression is not trivial. Workaround is using a normal distribution with the deterministic output as mu and a small sd (.1 - 1) and observed=data.
So you can first do s_data = theano.shared(data), and then after you write down your model, you use theano.tensor.stack or theano.tensor.concatenate to concatenate the latent observed with s_data and feed it to MvGuassianRandomWalk.