Enhancing classifier performance using an improvised ensemble model: An application to credit scoring

Abstract

Credit scoring is a mathematical approach that many financial organisations use as the basis for determining whether to offer consumers financial products such as loans after carefully reviewing their transaction history. Typically, this is done to lower costs for the firm and eliminate any risk hazards that occur due to default on the extended loans. This is usually done through various classification techniques which divide the customers into various buckets based on their past data. With just 1% increase in the classification accuracy, companies can observe better profitability. We have performed several individual techniques like logistic regression, decision trees, neural network etc. However, these individual classification methods are often less accurate and have assumptions and bias involved. Ensemble methods on the other hand, correct this bias and produce a highly accurate model. This involves aggregating different individual classifiers and assign different weights to their scoring based on relative accuracy. Thus, this project develops different ensemble based models and compares them on various metrics to identify the one that performs well on the chosen datasets. The two datasets chosen are Australian and German datasets having different number of feature variables. As a result, on both these datasets, there was a statistically significant improvement in performance versus the best performing individual models. On Australian dataset, the best achieved accuracy was 89% and on German dataset it was 88%. Weighted soft voting ensemble performs better than other individual and ensemble based techniques. This project is especially crucial for modem fin-tech companies that can create customised products based on credit scores as well as for the traditional banks that offer loans to their consumers.