A set of flexible, efficient and easy to use functions to build a credit scoring model from beginning to end, including dataset reading and preprocessing, variable binning and woe-encoding, model performance and stability validation, as well as visualization.
install.packages("devtools")
devtools::install_version("smbinning", version = 0.4)
devtools::install_github("xxzcool/scoremodel")This package depends on 'smbinning' prior to version 0.4, so you have to install the specified version of 'smbinning' at first.
The following is a demo for illustrating how to use this package to establish a credit scoring model from beginning to end. You can run the following codes manually, or use demo(scoremodel,package="scoremodel") directly.
library(scoremodel)
data(CreditData)
#convert character variables to factors
mysample <- convertType(CreditData, toType = "fac", vars = -1)
#homogeneity exclusion
mysample <- delNArate(mysample, narate_critical = 0.9)[[3]]
mysample <- delFewValues(mysample, minN = 5, minF = 2, exclude = "target")
mysample <- delSinvalPercent(mysample, percent = 0.9, exclude = "target")
#split dataset to train and test
set.seed(21)
splitresult <- splitData(mysample, size = 0.7, ifpercent = TRUE)
train <- splitresult[[1]]
test <- splitresult[[2]]
#optimal binning and generate woe configuration list
configlist <- dfBinningFun(train, binMethod = 1, p = 0.05, maxcat = 10, aliquots = 5)
#woe encoding
x_train_encoding <- woeEncodeFun_df(train[-length(train)], config_list = configlist)
x_test_encoding <- woeEncodeFun_df(test[-length(test)], config_list = configlist)
train_encoding <- cbind(x_train_encoding, train["target"])
#collinear elimination
train_encoding_red <- collElimination(train_encoding, cor_critical = 0.8)
#model training
fit <- LRfit(train_encoding_red, sig = 0.05)
summary(fit)
#prediction
p_train <- LRpredict(fit)
p_test <- LRpredict(fit, newdata = x_test_encoding)
#calculate KS of model
myks(train[,"target"], p_train)
myks(test[,"target"], p_test)
#calculate PSI of model
psi(p_train, p_test, bins = 10, binMethod = "EF")
#model performance visualization
myCurves(train[,"target"], p_train, test[,"target"], p_test, ontest = FALSE,
lift_bins = 10, P0 = 600, PDO = 50, color_scheme = 1, ifsave = FALSE)
myCurves(train[,"target"], p_train, test[,"target"], p_test, ontest = TRUE,
lift_bins = 10, P0 = 600, PDO = 50, color_scheme = 1, ifsave = FALSE)- Fix bug of 'preBinningFun': if font file 'msyh.ttf' does not exist in current system, then try to use 'msyh.ttc' file.
- Add a demo to the package.
- Fix bug of 'convertType': check beforehand if there is at least one variable which can be converted to avoid errors.
- Fix bug of 'LRfit': adaptable update with dependency package 'broom' to avoid errors.
- Speed up four functions, including 'woeEncodeFun_df', 'executeBinFun_df', 'rawPredictFun_df' and 'crossValidation'.
- No longer export four less-used auxiliary functions, including 'woeEncodeFun', 'executeBinFun', 'rawPredictFun', 'maxSinvalPercent_x'.
- Greatly speed up three functions, including 'delSinvalPercent', 'dfIV' and 'collElimination'.
- Fix bug of 'LRfit': add recursive judgment to filter out insignificant variables completely.
- Add an auxiliary function of 'insertElement'.
- Fix bug of 'psi': add two parameters of bins number and binning method, and set its percentage to a minima if the bin is missing to avoid calculation failure.
- More flexible of 'convertType': rename function from 'convertToFactor' to 'convertType', and add two parameters of toType and vars to integrate three types of conversions and be able to specify converted variables.
- Rename some output files of 'preBinningFun'.
- Travis CI building pass for the first time.
- First release.