Ordering of bullet points indicates (rough) chronological order of implementation.

Models

  • More GAS distributions – Poisson/Binomial/Gamma/Exponential distributions
  • GAS optimization fix – optimization is currently very unreliable for GAS models; intend to implement a custom optimization routine.
  • Explanatory variable inclusion for ARIMA/GAS
  • VAR models
  • Linear/gaussian state space models
  • Gaussian process AR/GAS models
  • Kernel AR/GAS models
  • RNNs

Inference

  • Hamiltonian Monte Carlo – skeleton of code is complete; need to implement automatic differentiation.
  • Black Box Variational Inference
  • INLA
  • SGD & AdaGrad
  • Expectation Propagation
  • Simulation samplers for state space models
  • NUTS-HMC
  • SG-Langevin

Optimization/speed-ups

  • Laplace approximation – use analytical pdf for graphs.
  • Prediction intervals – include model uncertainty if using Bayesian inference.
  • Conditional least squares option for ARMA models