Unobserved components model using the linear innovations state space representation (single source of error).
Key features:
- Estimation using autodiff (TMB)
- Model selection and ensembling
- Regressors in the observation equation (non-time varying)
- Choice of distributions (Gaussian, Student and Johnson’s SU)
- Choice of dynamics in variance (constant or GARCH)
Methods for specification (modelspec), estimation (estimate),
summary with choice of vcov (summary and vcov), diagnostics
(tsdiagnose), online filtering (tsfilter), prediction (predict),
simulation (simulate), backtesting (tsbacktest) and profiling
(tsprofile). Additionally, model selection (auto_select) and
ensembling (tsensemble) is also included.
The default algorithm from the nloptr solver is SLSQP which will sometimes return NaN in the sampled parameters during a difficult optimization. This will immediately terminate the estimation with an error. The termination was introduced in version 1.0.2 of the package else it would crash the R process as a result of a crash in the underlying RTMB eigen evaluation function.
Currently, we try to catch these cases whenever there is no default
control list passed (i.e. it is NULL) and switch to another algorithm in
those cases (see documentation of estimate). However, for safety, the
users can wrap the estimate in a try expression, and when it does fail
can switch to using the AUGLAG/MMA algorithm of the nloptr solver which
does not return NaN parameters:
# default algorithm is SLSQP
mod <- try(estimate(spec), silent = TRUE)
if (inherits(mod, 'try-error')) {
mod <- estimate(spec, control = issm_control(algorithm = "AUGLAG/MMA"))
}