Filtering vs MCMC

[hyunjimoon]

@jandraor and I discussed as follows, but due to the power of convergence diagnostics we will live in MCMC realm for some time.

• Formulation of the model affects the estimating method you use
• Jair thinks: particle filtering might be preferred to MCMC when more flexible
• Angie thinks: i) demand: we need shorter time scale and online decision, ii) supply: time-series data
  ParticleFiltering_recommendByNathanielOsgood.pdf
  

Q. sequential monte carlo (= particle filter) vs sequential Bayesian update we think the same

[hyunjimoon]

Summary of 879 seminar

• Kalman filtering is very black-boxed (decouples flow computation and stock-flow structure), so several safety nets are needed e.g. defining additional stock variable S' = max(S, 0) and connecting it to the next variables.
• process noise (compared to white noise) is easier when stochasticity and filtering comes in, as it is modularized stochasticity (recall stanify's function structure here which returns process_noise_dydt

Hazhir's assignment, I wish to discuss the following parts:

With process noise oscillations will be continuously elicited by the random shocks, and the system will continue to oscillate indefinitely, as if it was undamped, albeit with some irregularity in the periodicity of oscillatory modes.

state-resetting is more data-hungry!

wouldn't close-following data requires less data? State-resetting seems to have the danger of overfitting.