A Bank-Account-Information-Based Credit Scoring Method
with Bayesian Hierarchical Modeling

Overview

With the cooperation of Johoku Shinkin Bank, Emerada conducted joint research with Associate Professor Rei Yamamoto (Faculty of Science and Technology, Keio University) and Associate Professor Suguru Yamanaka (College of Science and Engineering, Aoyama Gakuin University) on building a bankruptcy-prediction model that uses corporate transaction information — bank account balances and deposit/withdrawal records. The resulting paper was accepted by the International Journal of Financial Engineering (World Scientific).

The study segments target firms by annual sales size and firm age, and examines how explanatory variables derived from transaction information affect the prediction of future bankruptcy.

Method and results

The degree to which account information predicts bankruptcy differs by a firm’s sales size and age. The study captures this segment-level heterogeneity with a Bayesian hierarchical model, which also mitigates the data-sparsity problem introduced by segmentation.

Empirically, the resulting account-information-based scoring model outperforms the traditional logistic regression model in predictive accuracy. The paper was accepted on the strength of the novelty of both the data used and the prediction method.

Practical significance

In practice, the model makes it possible to assess credit risk using account information for younger firms with little financial track record — precisely the firms for which conventional financial-statement-based scoring performs poorly.

Citation

Please refer to the journal page for the authoritative author list, volume, and pages.