Loan Interest Rate Regression Analysis (R)

Overview

This project uses historical personal loan data from 10,000 borrowers to develop and evaluate regression models for predicting loan interest rates. The analysis includes data preparation, feature engineering, train-test validation, correlation analysis, model comparison, and evaluation on unseen data.

The project was originally developed as part of graduate-level data science coursework and was subsequently polished for inclusion in my data analytics portfolio.

Objective

The objective of the analysis is to identify a regression model that effectively predicts personal loan interest rates using borrower and loan characteristics.

Four candidate linear regression models were developed and compared using Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and adjusted R-squared.

Data Preparation

The original dataset contains 10,000 borrower records. Relevant variables were selected and renamed for analysis.

Data preparation included:

• Examining the dataset for missing values

• Investigating missing debt-to-income ratios

• Replacing missing individual debt-to-income ratios with corresponding joint debt-to-income values where appropriate

• Converting loan issue dates to month

• Creating binary indicators for loan term and selected loan purposes

• Converting categorical predictors to factors

• Removing redundant source variables following transformation

The cleaned dataset was then divided into an 80% training set and 20% test set using a fixed random seed for reproducibility.

Exploratory Analysis

Correlation matrices were generated for the training and test datasets to assess relationships among numeric variables and determine whether those relationships were reasonably consistent between the two samples.

Regression Modeling

Four candidate regression models were evaluated using different combinations of borrower and loan characteristics.

The models incorporated predictors including:

• Annual income

• Loan term

• Credit utilization

• Loan grade

• Issue month

• Debt-to-income ratio

• Recent credit inquiries

• Loan purpose

Interaction effects between income, credit utilization, and loan term were also evaluated in one candidate model.

Model Comparison

The four models were compared using:

• AIC — lower values indicate a better relative fit

• BIC — lower values indicate a better relative fit while imposing a stronger penalty for model complexity

• Adjusted R-squared — higher values indicate greater explanatory power while accounting for the number of predictors

The model using loan grade as the sole predictor performed best across all three criteria.

Metric

Model 3 Result

AIC

24,170.03

BIC

24,225.92

Adjusted R²

0.952

Despite its simplicity, Model 3 outperformed the more complex candidate models.

Test-Set Performance

Model 3 was evaluated on the previously unseen test dataset.

The model achieved an adjusted R-squared of 0.947 on the test data, compared with 0.952 on the training data. The similar performance across the two datasets indicates that the model generalized well to unseen observations.

Key Finding

Among the candidate models evaluated, loan grade alone provided the strongest predictor of interest rate. More complex specifications incorporating borrower income, debt-to-income ratio, credit utilization, loan purpose, loan term, and interaction effects did not improve overall model performance.

The results also demonstrate the value of evaluating model complexity rather than assuming that adding predictors necessarily produces a better model.

Tools and Libraries

The analysis was completed in R using:

• tidyverse

• car

• lm.beta

• olsrr

• nnet