Long waiting for a credit decision is a pain, which is indicated by a large proportion of bank customers.
Long waiting for a credit decision is a pain, which is indicated by a large proportion of bank customers. The formulation of an application, assessment of creditworthiness, collateral, risks, analysis, opinion and acceptance are the process steps necessary for granting a credit from the bank’s point of view. The decision-making process may look a bit different if the borrower is a client serviced by a financial institution for a long time. In such cases, the lender has a great deal of knowledge about the history of cooperation – it is possible not only to speed up the process of application verification, but even to prepare a credit offer without an application from the client. The article presents the implementation of the module supporting the generation of individual credit offers using Machine Learning.
Whiteaster’s task was to prepare the automation of the creditworthiness assessment of a financial company’s client. The company’s clients are merchants who sell through one of the world’s largest Internet sales platforms, Amazon. The credit offer was to be based on creditworthiness and risk analysis. It assumes that the credit should be possible to be repaid thanks to the sales results on Amazon. These loans are a new product – the company did not have data on the history of their granting and repayment.
The task is divided into several steps:
Since in the first stage the offer is addressed to current customers (using other financial products), the company has a large amount of data. They are divided into two main groups:
The historical data relate to sales, customer service quality and status on the Amazon platform. The available sets of information, apart from the fact that they differ in substance, cover different time periods (from two to six months).
The second set is current data describing the activity of the Amazon salesmen – current status information, information on sales, products offered, financial and accounting data, information on loans granted (currently repaid), etc.
Analysis of available information and development of a decision-making model is a key element of the project. In practice , it is a continuous task performed all the time during the project implementation. Obtained indirect results , conclusions, diagnosed special cases, revealed in the course of work and growing knowledge in the team result in the evolution of models, methods and the list of data sets in the project. As a result, they lead to system improvement. The methods of data classification are used intensively, which facilitate the discovery of regularities and relationships. The knowledge and experience of the client’s specialists cannot be overestimated. Good cooperation, jointly developed solutions and ongoing evaluation of results allow for quick progress.
The analytical model was based on the conclusions from the analysis of data and its impact on credit decisions. The final loan offer results from many factors, both of historical and current nature. One may risk a statement that the observed increase in their quantity is a good illustration of the maturation of decision-making systems. The architecture of the solution is also affected by the method of data acquisition, the frequency of its updating and the substantive meaning – the way it affects the final recommendation. As the results of the analysis change – they evolve during the project, similarly, changes in the model follow. As a result of the project, a solution consisting of seven Machine Learning models and a calculation module based on designed algorithms was created.
The decision-making system generates a ranking (Scoring beha w ioralny , corresponding to the degree of credit risk) and a loan amount proposal for each client. It also provides a range of supporting information to facilitate verification but also a justification for the decision. The architecture consists of seven Machine Learning modules in total – five of them analyse historical data. Different models available in the O.ai package are used. Additionally, a number of rules and algorithms constituting the calculation module have been implemented. Teaching ML modules and improving algorithms is a time-consuming process. It requires many tests, changes, repetitions, result analyses, special case diagnoses, etc. The solution accepted for production use – based on its recommendations of credit decisions – was created in less than 1.5 months. Integration with the client’s integrated system took another month.
The project implementation required the knowledge of the information model, data model and business model of the Client’s company (not only in relation to the product that was the subject of the works). Good cooperation and highly evaluated results made it possible to define further goals and continue the cooperation.