Title: Delayed payment modelling using machine learning methods Authors:  Mindaugas Kavaliauskas - Kaunas University of Technology (Lithuania) [presenting]
Abstract: Credit risk modelling holds particular significance for trade companies because it's common for them to permit the purchase of goods with deferred payment terms. Risk assessment models typically estimate a company's risk by using various financial metrics such as net profit, total revenue, working capital, total assets, and their corresponding ratios. One example of such a model is the Altman Z-score, a classic bankruptcy prediction model first published in 1968. However, these models have a significant drawback. These models are not well-suited for real-time risk assessment. Financial reports are typically published months after the end of the financial year. The aim is to utilize an alternative data - delayed invoice payment times for credit risk assessment. The attempt is to forecast future payment delays using past payment delay records. While this may seem like a typical time series forecasting problem, the data's nature is quite distinct: the time intervals between invoices are irregular, and the number of invoices for a particular company can vary from just one invoice in its history to over a dozen invoices per month. Building models based on this kind of data requires additional data preprocessing procedures such as padding, trimming, etc. Several data preprocessing methods are explored, and a few statistical and machine learning models are applied. The accuracy of these models is provided and discussed.