knitr::opts_chunk$set( collapse = TRUE, comment = "#>" )
After installing scorecard via instructions in the README section, load the package into your environment.
library(scorecard)
Let's use the germancredit dataset for the purposes of this demonstration.
data("germancredit") str(germancredit)
The var_filter
function drops column variables that don't meet the thresholds for missing rate (> 95% by default), information value (IV) (< 0.02 by default), or identical value rate (> 95% by default).
dt_f <- var_filter(germancredit, y = "creditability")
When building scorecard models, a subset of the observations should be held out from the data used to train the model (similar to most other traditional modeling approaches), and instead be apportioned to the test set. We can perform this sampling to create the train and test datasets using the split_df
function.
dt_list <- split_df(dt_f, y = "creditability", ratios = c(0.6, 0.4), seed = 30) label_list <- lapply(dt_list, function(x) x$creditability)
Weight-of-Evidence binning is a technique for binning both continuous and categorical independent variables in a way that provides the most robust bifurcation of the data against the dependent variable. This technique can be easily executed across all independent variables using the woebin
function.
bins <- woebin(dt_f, y = "creditability") # woebin_plot(bins)
The user can also adjust bin breaks interactively by using the woebin_adj
function.
# breaks_adj <- woebin_adj(dt_f, y = "creditability", bins = bins)
Furthermore, the user can set the bin breaks manually via the breaks_list = list()
argument in the woebin
function. Note the use of %,% as a separator to create a single bin from two classes in a categorical independent variable.
breaks_adj <- list( age.in.years = c(26, 35, 40), other.debtors.or.guarantors = c("none", "co-applicant%,%guarantor") ) bins_adj <- woebin(dt_f, y = "creditability", breaks_list = breaks_adj)
Once your WoE bins are established for all desired independent variables, apply the binning logic to the training and test datasets.
dt_woe_list <- lapply(dt_list, function(x) woebin_ply(x, bins_adj))
Logistic regression can often be leveraged effectively to assist in building the scorecards.
m1 <- glm( creditability ~ ., family = binomial(), data = dt_woe_list$train) # vif(m1, merge_coef = TRUE) # summary(m1) # Select a formula-based model by AIC (or by LASSO for large dataset) m_step <- step(m1, direction = "both", trace = FALSE) m2 <- eval(m_step$call) # vif(m2, merge_coef = TRUE) # summary(m2)
If oversampling is a concern, the following code chunk could be uncommented and run to help adjust for this issue.
# Read documentation on handling oversampling (support.sas.com/kb/22/601.html) # library(data.table) # p1 <- 0.03 # bad probability in population # r1 <- 0.3 # bad probability in sample dataset # dt_woe <- copy(dt_woe_list$train)[, weight := ifelse(creditability == 1, p1/r1, (1-p1)/(1-r1) )][] # fmla <- as.formula(paste("creditability ~", paste(names(coef(m2))[-1], collapse = "+"))) # m3 <- glm(fmla, family = binomial(), data = dt_woe, weights = weight)
The perf_eva
function provides model accuracy statistics (such as mse, rmse, logloss, r2, ks, auc, gini) and plots (such as ks, lift, gain, roc, lz, pr, f1, density).
# First, get probabalistic predictions pred_list <- lapply(dt_woe_list, function(x) predict(m2, x, type = 'response')) # Then evaluate model accuracy perf <- perf_eva(pred = pred_list, label = label_list)
Once the model has been selected, scorecards can be created via the scorecard
function. Note that the default target points is 600, target odds is 1/19 and points to double the odds is 50. See ?scorecard
for more information on the function and its arguments.
The scorecard can then be applied to the original data using the scorecard_ply
function. Lastly, a chart encompassing Population Stability Index (PSI) statistics can be rendered via the perf_psi
function.
# Build the card card <- scorecard(bins_adj, m2) # Obtain Credit Scores score_list <- lapply(dt_list, function(x) scorecard_ply(x, card)) # Analyze the PSI perf_psi(score = score_list, label = label_list)
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.