mmsrank_int: Rank models based on leave-n-out (LNO) score

In reumandc/mms: Model Selection Via Resampling for Matrix Models

Description

Function to rank models using a leave-n-out cross validation (LNOCV) procedure for matrix regression models. Same as mmsrank, but with no error checking.

Usage

mmsrank_int(mats, model.names, n, maxruns, rank.mod = F)

Arguments

mats	A list of matrices, all assumed to be the same dimensions. Only the lower triangles are used. NA/NaNs are allowed. The first entry taken to be the response.
model.names	A list of models to run LNOCV on. NA input not accepted, in contrast to mmsrank. Specification needs to be as numeric values that correspond to mats elements. Examples of model specifications: 2, 2:3, c(2,3,5).
n	The number of sampling locations to leave out. Must be at least 2.
maxruns	The maximum number of leave-n-outs to do - to be used if choose(dim(mats[[1]]),n) is very large. Inf to use (or try to use) all LNOs. If maxruns is a number, then LNOs are selected randomly and hence may include repeats.
rank.mod	Logical. If TRUE, sort models by rank. If FALSE (default), do not rank models.

Value

mmsrank Return a data frame with columns for

model.names	The name of the model, based on the indices of included predictors in mats
lno.score	The out-of-sample forecast accuracy (mean squared error)
num.pos	The possible number of LNOs for the given n and number of locations
num.att	The total number of LNOs attempted
num.rnk	The number of LNOs that did not result in a rank deficiency regression problem, and so could be used for testing out-of-sample predictions
num.usd	The number of LNOs that could be used in the end (possibly less than num.rnk because of NAs in the input matrices)

Author(s)

Tom Anderson, [e-mail]; Daniel Reuman, [e-mail]; Jon Walter, [e-mail]

reumandc/mms documentation built on May 28, 2019, 5:39 p.m.