Credit Risk Modeling using SAS

Set historical window, observation point, and performance window to design a behavior scorecard, gathering independent variables and good or bad outcomes to define the target and align with Basel guidance.

Exclude fraudulent accounts, staff and VIP loans, out-of-country and pre-approved loans, lost or stolen cards, underage or voluntarily closed accounts, and markets no longer operating from the development sample.

Explore the retail dataset in SAS Studio, examine 23 variables across 18,267 observations, and understand how demographics, financials, payment behavior, and debt ratios relate to default.

Identify factors for selecting historical and performance windows in credit risk modeling, including seasonality, model type, and vintage analysis, with typical 12–24 month and 6–12 month windows.

Develop a behavioral scorecard using past payment behavior, demographic information, and credit bureau status to predict one-year default risk with logistic regression for existing personal loan customers.

Explore sample data fields used to build a retail portfolio scorecard, including demographics, financials, collateral, credit bureau status, and payment behavior to predict loan default.

Explore vintage analysis, aka cohort analysis, to determine the performance window by tracking cumulative bad rate across months on books, revealing 24 months as the stabilization point.

Roll rate analysis uses delinquency buckets to define bad customers as 90 days or more past due, via observed roll-back and roll-forward transitions.

Logistic regression uses a binary logit model for scorecard development, handling a 0/1 dependent variable with flexible independent variables and estimating the default probability (Y=1) as a score.

Clean data for credit risk modeling by identifying missing values and outliers with SAS. Use proc means, proc freq, and proc univariate to summarize numeric and character variables.

This lecture demonstrates identifying missing values and outliers with univariate analysis using proc means, and removes them with proc sql to create a clean retail one dataset for future modeling.

Use information value to screen variables by their individual predictive strength for binary logistic regression, dropping low-IV predictors (<0.02) and guiding reduction before multicollinearity assessment and final model building.

Explore fine classing and coarse classing to group continuous and categorical variables. Understand how these non-linear patterns improve credit risk modeling in SAS.

Explore fine and coarse classing for credit risk modeling, using deciles, weight of evidence (W0), and continuity to create bins and smoother nonlinear risk curves.

Calculate the weight of evidence for each bin after coarse classing, create income_W0 for modeling, and use information value to screen variables with IV ranges guiding predictive strength before multicollinearity.

Develop WOE and IV values in SAS using a bivariate macro to compute default totals and rates, then interpret and screen variables for credit risk modeling.

Learn to perform coarse classing in SAS Studio, create continuous and categorical bands, compute IV and WOE values, and generate WOE variables for credit risk modeling.

Analyze multicollinearity using variance inflation factor thresholds and p-values in SAS with proc reg and proc logistic. Identify significant predictors like income, housing, age, and employment for default flag modeling.

Explore how to assess a binary logistic regression model in SAS using concordance, discordance, and tied pairs, and learn to interpret predicted probabilities (p hat) and model discrimination.

Explore Somers' D and C statistics from a logistic model output, computing Somers D as (percent concordant minus percent discordant)/100, with a C statistic of 0.775 indicating acceptable discrimination.

Learn how to validate a logistic regression model through rank ordering across deciles, and interpret KS statistics, the Gini coefficient, and the Lorenz curve to assess discriminatory power.

Apply the Hosmer-Lemeshow goodness-of-fit test to assess whether observed defaults align with predicted probabilities across deciles, using SAS proc logistic with lac fit rescue.

Perform a clustering check to ensure the scorecard differentiates similar accounts, minimizing concentration at any score point (ideally 5–6%), using p hat scores and SAS proc freq.

Learn to validate a final logistic regression model using SAS proc logistic, comparing development and validation samples, checking parameter signs, p-values, and rank ordering through validation rerun and scoring.

Compute the Brier score to assess calibration by averaging the squared differences between predicted probabilities and actual outcomes across bins. A lower score signals a more accurate credit risk model.