Online Learning with State Space Linear Kalman Filter

[Dekermanjian]

It would be really cool to be able to update a Linear Kalman filter state space model as new data are available -- as opposed to re-fitting the entire model every time you get new data.

Below are the potential steps required to make this happen (quoting @jessegrabowski ):
"

1. Compile update and predict from the KalmanFilter class separately
2. Make some kind of database class to hold incoming data
3. Have the actual PyMC model set aside somewhere

I would have 3 workers running async:
Worker 1 produces predictions given the current state of the filter and hyperparameters every unit of time
Worker 2 listens to a data stream and updates the filter state when a new data point arrives
Worker 3 runs the full PyMC model and updates the hyperparameters every longer time interval
"