Predicting Loan Defaults

Aim of the project

Create a risk assessment tool to help LendingClub, an online P2P loan provider, to understand whether an applicant is likely to pay a loan back or fall into default.

This repository contains underlaying code and data sources that I used to build a machine learing model (binary classifier).

Click here to access the presentation.

To process data and to create the model I worked with the programming language Python.

This is my final project as part of the CodeClan's Data Analysis course training.

Structure of the project - predictive analysis, classification:

A. Pre-modeling (data processing)

1. import libraries

2. data import and exploration (summary statistics, vizualizations)

3. data cleaning:

   • drop features (columns) with more than 50% of values missing,
   • drop features that only get populated once a loan has been granted (as model only inputs data for a new applicant)
   • drop features containing just one constant value, drop categorical features showing high cardinality (= too many levels) that are not needed for machine learning and such as not worth further manipulation (eg. "binning")
   • drop observations (rows) with missing values "regex" operations on Python object values (eg. get rid of '%' symbol)
   • transform data types where needed (eg. str -> float)
   • drop observations related to current loans (as model only inputs data for a new applicant)

4. feauture engineering:

   • set target variable "loan_status" as binary (0 = default, 1 = paid)
   • categorical features "binning" - lower the cardinality by binning values into eg. quartile intervals (results into just 4 levels and data less affected by outliers)
   • apply Box-cox transformation to address skeweness (make variables distribution more normal)

5. feature selection:

   • train/test split first to avoid overfitting!
   • on train and test datasets separatelly:
     • perform further feature reduction
       • based on correlation (between features, and between features and target variable)
       • based on variation

6. feature dummying

   • dummy categorical variables (models can only handle numerical features)

B. Model building - machine learning binary classification model:

1. Logistic regression,

2. Gaussian Naive Bayes,

3. Random Forest

   • target variable: loan_status (binary variable)
   • feature matrix (explanatory variables): all features engineered and selected in the pre-modeling step
   • null model for cross validation
   • build models on default classification threshold 0.5 (Logistic regression, Gaussian Naive Bayes, Decision Tree, Random Forest)
   • access model performance
   • calculate metrics: accuracy, precision, sensitivity (= recall), specificity, AUC
   • visualize ROC
   • visualize sensitivity and specificity graphs by modifying the classification threshold and find the optimal threshold
   • compare the predictive power of adjusted models (TP and TN on confusion matrices) and choose the best model

C. Exploratory data analysis

• list out main features that have direct and clear impact on an applicant being likely to pay the loan or fall into default
• showcase model application on random sample of loan applicants

Python libraries

An example of used Python libraries that were fundamental for the project - used to read in, clean, transform, analyse and visulase data, and to train the models and evaluate their performance:

numpy, pandas, matplotlib.pyplot, seaborn, sklearn

Code example