I think there are two ways to add ARIMA to pymc3. The first is that we use statsmodels to implement arima in pymc3, and then continue to improve on it, such as automatic arima like pyramid.arima. The second is to implement an ARIMA ourselves.
We can’t use
statsmodel because it uses
numpy as the computational backend, whereas pymc3 uses
theano to be able to do symbolic math and automatic gradient calculations.