LendingClub (C): Gradient Boosting & Payoff Matrix

This case builds directly on the LendingClub (A) and (B) cases. In this case students follow Emily Figel as she builds an even more sophisticated model using the gradient boosted tree method to predict, with some probability, whether a borrower would repay or default on his loan.
Having now built three models, Figel compares them to determine which model is most effective at classifying borrowers correctly then uses that model to determine how to invest in a portfolio of loans. Students explore the relationship between her model’s confusion matrix, which organizes the model’s correct and incorrect classifications, the cutoff point on the curve that matches true positives and true negatives, and the payoff matrix Figel constructs. Students can then follow the link directly from the model Figel builds to her specific investment decisions.
Technical topics include: (1) Gradient boosted trees as a modelling technique; hyperparameters and learning rate; model validation; and (2) Evaluating model output; ROC curve, cutoff point, confusion matrix; payoff matrix as a framework for utilizing the model to compare investment opportunities.