binnr is an R package that helps modelers build scorecards. Scorecard models
need to be more than predictive. Regulatory oversight often necessitates they be
transparent as well. Model transparency is difficult to enforce in more
predictive, non-linear methods such as neural networks, random forest, or
gradient boosted decision trees. Often direct modeler intervention is required
to ensure proper treatment is given for values of independent variables.
binnr attempts to solve these problems by providing interactive variable
manipulation facilities giving the modeler total control over how variables
are treated in scorecard models.
if(!require(installr)) install.packages("installr")
installr::install.Rtools()
With all of the prerequisites installed, installing binnr is easy:
## used to install packages directly from gitlab
if(!require(devtools)) install.packages("devtools")
devtools::install_git("https://gitlab.ins.risk.regn.net/minneapolis-r-packages/binnr.git", build_vignettes=TRUE)
From within R, type the following to open a cheat sheet with binnr commands:
vignette("binnr-cheat-sheet")
Modeling
8000
with binnr follows four phases of development.
- Bin
- Fit
- Review
- Adjust
The model is usually binned just once, while steps 2 - 4 are repeated as many times as necessary.
The beta version of binnr differs drastically from previous versions in its implementation. Only one scorecard object is created for each modeling endeavor.
All of the operations performed on a Scorecard object are performed "in-place". This is different from how R typically works where the return value from functions has to be assigned to a variable or lost.
mod2 <- fit(mod1)
mod3 <- adjust(mod2)
mod$fit("model 1")
mod$adjust()
library(binnr)
data(titanic)
mod <- bin(data=titanic, y=titanic$Survived)
## optional arguments controlling binning
mod <- bin(
data = titanic,
y = titanic$Survived,
w = weight,
min.iv = 0.01,
min.cnt = 100,
min.res = 25,
max.bin = 10,
mono = 2,
exceptions = c(-1))
## check for multicollinearity
cc <- mod$cluster()
to_drop <- mod$prune_clusters(cc, 0.90, 1)
mod$fit("model 1", "Initial fit with all variables")
mod$sort()
mod$summary()
mod$compare("model 1", "model 2")
mod$select("model 1")
mod$get_dropped()
mod$get_inmodel()
## show and plot are called automatically from the adjust method
mod$variables$Pclass$show()
mod$variables$Pclass$plot()
mod$drop(c("Pclass", "Fare"))
mod$undrop(all=TRUE)
## no longer need to assign to an object
mod$adjust()
## old way of using binnr
# mod2 <- adjust(mod1)
## run 20 bootstrap samples
pvals <- mod$pseudo_pvals(times=20, bag.fraction=1, replace=TRUE)
## drop variables that have zero coefficients at least 10% of the time
to_drop <- names(which(pvals$pvals > 0.10))
mod$drop(to_drop)
mod$fit("final model")
## can calculate RCs from min/max/neutral
code <- mod$gen_code_sas(pfx="my_mod1", method="min")
cat(code, sep="\n", file="path/to/my/model.sas")
saveRDS(mod, "my_model.rds")
mod <- readRDS("my_model.rds")