Since structural time series models only require fitting process and observation error variance parameters (unlike ARIMA models, for example), the gradient of the log-likelihood function can be computed analytically, providing more efficient parameter estimation than gradient-free ML optimization or EM estimation (which requires repeated smoothing), the two techniques currently implemented here.
I have a very minimal local level + linear regression model with diffuse state initialization up and running at https://github.com/GordStephen/TinyTimeModels.jl - ultimately I think it would be very cool to have a modular structural time series model interface here that would allow for mixing and matching particular components (local level or linear trend random walks, seasonal trends, regression terms, etc).
milktrader
commented
8 years ago
Since structural time series models only require fitting process and observation error variance parameters (unlike ARIMA models, for example), the gradient of the log-likelihood function can be computed analytically, providing more efficient parameter estimation than gradient-free ML optimization or EM estimation (which requires repeated smoothing), the two techniques currently implemented here.
I have a very minimal local level + linear regression model with diffuse state initialization up and running at https://github.com/GordStephen/TinyTimeModels.jl - ultimately I think it would be very cool to have a modular structural time series model interface here that would allow for mixing and matching particular components (local level or linear trend random walks, seasonal trends, regression terms, etc).