Predicting Vehicle Loan Default using AdaBoost

This project addresses the challenge of predicting vehicle loan default of Indian borrowers using data acquired in 2019 from the LTFS Data Science FinHack (ML Hackathon). Check out the link to the competition for a detailed dataset and problem description.

Project Overview

The objective is to accurately predict the probability of loanee / borrower defaulting on a vehicle loan in the first EMI (Equated Monthly Installments) on the due date. Python with Jupyter notebooks are used for the implementation, relying heavily on scikit-learn library. AUC is utilized as the evaluation metric due to class imbalance. In the project pipelines are extensively used to avoid data leakage and simplify transformations.

AdaBoost is chosen because it's a more sophisticated model capable of capturing non-linear patterns and uses less RAM compared to RandomForest.

Data Description

• data_dictionary.csv contains a brief description on each variable provided in the training and test set.
• train.csv contains the training data with details on loan as described in the last section
• test.csv contains details of all customers and loans for which the participants are to submit probability of default.

Project Structure

• The key emphasis in the project is modularity, which is why multiple notebooks and Python files are created. The data is first explored and investigated in detail in data_exploring.ipynb, and using the insights gained, modelling is then performed in modelling.ipynb. Transformations done in data_exploring.ipynb are extracted into functions in helpers.py so they can be effectively used in a pipeline.
• config.py contains global constants, denoted with uppercase letters, used in all notebooks.
• helpers.py contains custom functions and classes used in all notebooks.

Modelling

The key steps performed in modelling are following:

• Basic model training -> Evaluating performance of relatively basic AdaBoost models, one trained on a 50,000-sample subset and the other on the entire training dataset. The same test data is used for both models;
• Learning curves -> Creating two learning curves: one for the training set size and another for the n_estimators hyperparameter of AdaBoost;
• Tuning all hyperparameters except the split criterion for the AdaBoost estimator using BayesSearchCV;
• Tuning the 'criterion' hyperparameter for the base estimator of AdaBoost using GridSearchCV;
• Evaluating the model with the best hyperparameters using the default decision threshold on test data;
• Optimizing the decision threshold using the F1 metric;
• Evaluating the model with the best hyperparameters using the optimized decision threshold on the test data;
• Training the best model on the entire dataset;
• Evaluating the best model on the entire dataset;
• Creating predictions for test.csv and submitting results.

Results

Below are the AUC results for CV and the optimized AdaBoost model:

Dataset	AUC
CV	65.888%
Train Set	67.144%
Test Set	66.302%
Entire Set	67.069%

This approach would approximately yield a ranking of 113th place on the Private Leaderboard.

Presentation

The presentation in English is available as PDF in this repository: link.