Machine Learning and Finance Model in Predicting Default: Merton-based Reasoning

Abstract

When economics is combined with artificial intelligence in default prediction, difficulty arises because of the basic methodological gap, theory, and technical learning. This research empirically explores the difference between the Merton model, which calculates the probability of default by comparing the expected assets to its projection, and case-based reasoning (CBR), which learns microdata patterns. The research aims to define the characteristics of each method through a series of experiments. In addition, this study's goal is to synthesize a combined method that utilizes the advantage of each approach. Results show that the Merton model is more accurate than the CBR under a particular subjective condition. Few studies have explored and compared experimental studies between economic approaches and machine learning. The novelty of this research is how to form an optimum synthetic method of those different approaches. We recommend combining these models in the proposed algorithm to optimize the prediction performance. Furthermore, this work provides evidence supporting the fusion that yields 84% accuracy and 8% Type II error.