Peer To Peer Social Lending Default Prediction With Convolutional Neural Networks

Abstract

Social lending or peer to peer lending (P2P lending) is a social digital lending marketplace where borrowers and lenders are interlinked. With the emergency of coronavirus medical risks and travel limitations, this has come as a lucrative business. However with huge industry lay-offs and small companies going under, the verification of a potential borrower's creditworthiness is important. The objective of peer to peer lending was to eliminate middlemen or banks and create a personal touch of the borrower and lender, but results have shown that those running away from traditional banks do so because of unpaid loans, bad debts or other slow payment behaviour. Information asymmetry, lack of demand and money supply, poor income verification, inaccurate credit reports and employment details, high debt to income ratios as well as inadequate anti fraud systems are some of the challenges in peer to peer lending. To mitigate these myriad of problems, machine learning algorithms are used to predict default, slow payments and fraudulent application behaviour. The measurement by classification of such creditworthiness of peer to peer lending marketplaces is done with new frameworks like deep learning convolution neural networks, XGBoost, CatBoost and LightGBM. These new class of classifiers performed better than the c4.5 decision tree, ensemble voting classifier, random forest and k nearest neighbor classifiers. The deep learning algorithms were also optimized for overfitting leading to even better performance. The implementation was done using Weka, Keras and Tensorflow. The datasets used include Prosper and Lending Club peer to peer datasets managed by their respective online lending houses. The study also finds that purpose, employment status, age and credit grade or prosper score are the best predictors for defaulting of loans.