Version 0.2.0 of PyFlux has been released on PyPi.
- Non-Gaussian state space models (examples soon)
- t-distributed GAS model
- Simulation smoother for state space models
- Significant speed-ups for all existing models in the range 1.3-1.5x
- Smoothed estimates now produced for Gaussian state space models
- Bug with GARCH forecasting fixed
- VAR likelihood rewritten – now faster for non-OLS estimation methods
- New predict/predict_is/plot_predict/plot_predict_is methods
- Code refactored and cleaned up
- Numerous other changes
I produce new documentation tomorrow. The next couple of releases are going to focus on performance for the computationally intensive models – in particularly, the GP-NARX and non-Gaussian state space models. I need, for example, to rewrite the M-H sampler for the latter because it is too slow at present for non-exponential distributed variables.