search

pymc-devs / pymc

Bayesian Modeling and Probabilistic Programming in Python
https://docs.pymc.io/
Other
8.72k stars 2.02k forks source link

Enabling vector autoregressions #4665

Open ckrapu opened 3 years ago

ckrapu commented 3 years ago

The AR distribution appears to be nearly complete for usage as a true vector autoregression parameterized by p cross-series coefficients, each of shape (d,d). The main change that has to be enacted is to use a dot product instead of elementwise multiplication here. However, I am unable to determine the role of the constant argument and why it necessitates the calculation of eps = value[self.p :] - self.rho[0] - x where, under the AR / VAR model, eps is assumed to have a diagonal Normal distribution.

ricardoV94 commented 2 years ago

There is some code example of Vector Autoregression here: https://www.pymc-labs.io/blog-posts/bayesian-vector-autoregression/

However, I am unable to determine the role of the constant argument and why it necessitates the calculation of eps = value[self.p :] - self.rho[0] - x

The constant is just an intercept term: y = rho[0] + x, where x are the convolved values and lagged coefficients from rho[1:]. The AR was refactored to V4 in #5734