Ross Taylor

Prediction intervals should reflect model uncertainty

1

If the model is Bayesian, the intervals should account for the uncertainty in the parameters. Bootstrapping is an option for the frequentist approach.

Normalizing Nile data - BBVI works fine, but less well on non-normalized data. Need to investigate further.

Especially for state space models. Aim has always been to replace mean-field structure with something more appropriate, but in the meantime need to flag in the documentation and examples that...

Cythonization

8

Now that a core model set is available, it is time to look to optimize the code to allow for quicker estimation and scalability to larger datasets. **Model** - [x]...

Unit tests for models and inference

2

Related to https://github.com/RJT1990/pyflux/issues/10. Tests should be done on the GitHub side rather than locally. Steps: - Set Up Travis (Issue #10 ) - Upload tests to a /tests folder within...

Would be nice to animate the learning process for dynamic regression coefficients.

First state defaults to 0 - should be smoothed out; is causing BBVI to nan-out with some datasets.

Need to random search this in preoptimization stage. Noticed that a DAR model didn't optimize well today because of inappropriate starting values.

Makes a hella lot of sense to use this instead of Monte Carlo gradients. Difficulty is that we're not using a automatic differentiation backend or utility, so would need to...

LLT and DynReg intervals do not work (undefined over some intervals of time) - needs fixing