Estimating borrower default probability using industry affiliation and corporate ties

Industry affiliation is one of the important factors in assessing a сredit risk. In the Russian Federation, OKVED codes are used to indicate the type of economic activity. However the code is declared by the company itself and is not checked against its actual activity, and some borrowers use this to present a high-risk activity as a low-risk one. In such cases, classical credit scoring models are forced to rely on inaccurate data when assessing risk, which may ultimately lead to financial losses. Moreover, such models usually treat a counterparty as an isolated object and disregard its connections with other companies, although these connections carry information about the company's actual position in the economy.

The goal is to develop a methodology for constructing vector representations of the industry affiliation of companies that depend not only on the declared OKVED codes but also on their position in the classification and on the company's connections, and to test whether such a representation improves the quality of borrower default probability assessment compared with the declared code.

The input data are described by a model that combines the graph of connections between companies and the graph of OKVED codes, in which the classification tree is extended with links observed through companies. A methodology for constructing vector representations of industry affiliation on this model is proposed. It extends the author's previously published approaches with three elements: the weight of each classification level in the aggregation along the path to the primary code is computed from the company's neighbouкhood; additional codes are aggregated with weights depending on their agreement with the neighbourhood; the parameters of all components are found jointly by minimizing a criterion that combines the default prediction error and the link reconstruction error.

On a dataset of anonymized corporate loans, the representation built by the methodology increased the Gini coefficient from 0.5941 to 0.6125 compared with the best of the known encodings of the declared code, with the same other features and the same decision rule.

The methodology allows credit scoring systems to take into account not only the borrower's declared codes but also its neighbourhood, so that borrowers with the same declared code and different neighbourhoods receive different risk assessments.