tsensemble.tsissm.selection {tsissm} | R Documentation |
Ensembling of Models and Predictions
Description
Ensembles estimated, predicted and simulated objects arising from the automatic selection method.
Usage
## S3 method for class 'tsissm.selection'
tsensemble(object, weights = NULL, start = 1, ...)
## S3 method for class 'tsissm.selection_predict'
tsensemble(object, weights = NULL, start = 1, ...)
## S3 method for class 'tsissm.selection_simulate'
tsensemble(object, weights = NULL, start = 1, ...)
Arguments
object |
an object of class “tsissm.selection”, “tsissm.selection_predict” or “tsissm.selection_simulate”. |
weights |
a vector of weights equal to the number of models to be ensembled. |
start |
the index for the state decomposition (when all lambda equal). If 1, will return the predicted states else will return the lagged predicted states (which can be summed to return the predictive distribution). |
... |
not used. |
Value
For the estimated object, a list with the weighted fit and errors, whilst for the predicted and simulated objects a list with the ensemble predictions as a “tsmodel.predict” object. If the models were all estimated with the same Box Cox lambda, then the weighted state decomposition is also returned inside the “tsmodel.predict” object. Additionally, for the predicted object, the ensemble analytic mean is also returned.
Note
Only the size of the weights is checked (should be equal to number of selected models), but not checks are performed on the values of the weights or whether they sum to 1. This is left to the user. When lambda is 1, the sum of component will be off by 1 versus the weighted distribution since the Box Cox transform contains an offset (so it is to be expected). To replicate the value of the predicted distribution by summing the decomposed components, the argument “start” should be set to 0 to return the lagged predicted components.