One of our goals while creating the Akbank Lab Medium account was to share both advantages and difficulties of what we have experienced in our projects that we are working on. As Akbank innovation center, we both take part in innovative projects with our relevant colleagues within the bank, and aim to create common added value by bringing them with the innovation ecosystem together.
Our Innovative Project Work # 3: Using Machine Learning in Loan Processes
In our Medium article in September, we stated that advanced analytics studies are one of the areas that we focus on as Akbank Lab, and we shared information about our automated machine learning platform that we have recently implemented. In this article, we wanted to talk about a study in which machine learning technology is positioned in a different field in line with our strategic priorities of the bank.
The usage areas of machine learning in banking are quite wide. By using this technology in every field and process where data is produced, it is possible to make sense of the data and produce more effective results, and to continuously improve the results you produce. Artificial intelligence technologies, including machine learning, are estimated to potentially generate $1 trillion additional value each year for the global banking industry, and it is thought that it will affect almost all areas from marketing and sales to risk management, from human resources processes to information technology applications [1]. While the result is providing the right solution to the right customer in the field of marketing, in the field of credit it can be seen as setting the most appropriate limits for customers and addressing customer risk classification more accurately. In order for machine learning to be efficient in different areas of use, we think that it is very important to find qualified data, to have a modern infrastructure and to establish this competence as a culture inside the bank. Thus, with the use of this technology, it becomes possible to create more accurate decision-making processes, increase in efficiency, less error making and customer experience improvement.
Credit and related processes, which offer new customer acquisition opportunities for banks and therefore one of the important income items, are also one of the most suitable areas for the use of machine learning. There are areas of use of machine learning at different stages of credit processes such as credit assessment, identity verification, and fraud prevention.
In recent years, we can see that non-bank players have also offered solutions in the field of credit, not just payment which is generally the first product that comes to mind, especially in Europe, Asia and the USA. While platforms like Amazon or Alibaba offered these solutions as a complement to their core business, companies such as Funding Circle started taking a role in the industry by focusing entirely on the loan business. In addition, we started to see that product diversity increased with different models such as crowdfunding and some innovative solutions were created according to the characteristics and the needs of different customer segments. In other words, credit, which is one of the most traditional products of banks, is an area where players compete besides banks and financial services.
In an area where competition has increased so much, it is much more important to differentiate and improve the way of doing your business. One of the important steps in credit processes is “developing credit models” and improvements in credit models have the potential to provide significant gains. Because, in most general terms, with more advanced models, loan applications can be evaluated according to risk situations with higher discriminatory power compared to traditional models, and how much loan to be given to which customer can be determined through the rules with decision support systems. Thus, by effectively managing the riskiness of the bank's loan portfolio, the portfolio quality is raised, and the loan rates that may become problematic can be brought to the optimum level as much as possible. There are 2 main components to increase the performance of credit models: improving analytical models and using new data. An example of the latter is the use of e-commerce shopping data. We see examples of this in different geographies. As Akbank, the project we will discuss here was the improvement of our analytical models, and we aimed to benefit from machine learning technology in the field of credit, too.
We conducted the project with our Risk Management, Credit Allocation and Technology teams and worked with a business partner that we believe in its competence and speed. We started the project 3 years ago, the use of machine learning in loan processes was not an experienced field for Akbank at that time, and we aimed to improve the efficiency of these processes by using the knowledge of an expert financial technology company.
Our Project Process
This project has been very instructive for all teams as it is our first project based on machine learning applications in the field of loans. It was one of the most critical stages to prepare test data precisely according to data security standards in order to quickly test the competencies of our business partner and its machine learning platform. In addition, it was critical that the data used in analytical models were “clean”. Cleaning and preparation of the data to be worked on took a significant amount of our time. In order to create and test machine learning models for the very large volumes of data prepared, we needed more advanced servers than the ones on which our existing models work, and new additional servers were established for this project.
We made our PoC project with the "backtesting" method using past data, and focused on calculating how Akbank could benefit if we had used these newly created machine learning models in past loan applications. The model we chose for PoC was the retail loan model and it was aimed to create more credit volume and improve the non-performing loan ratio. While creating the machine learning model, the R language was used and many more variables were produced compared to the existing models. The process of preparing the data, developing and evaluating the model took about 6 months, and as Akbank teams, we worked very closely with our business partner. After the PoC process, Python started to be used as a model development environment.
The project results were very promising. Comparing the current model performance with the project results, a significant increase was achieved in the model prediction power. While this increase was seen in all sub-segments of the model, low and high risk customers were better separated with higher accuracy rates. As a result, it was decided to invest in this technology and to expand its use in different models, realizing that the renewal of our current retail loan model with machine learning can provide significant returns for Akbank. In the following period, the processes of developing models for collection, credit card, micro and SME segments and products with various innovative modeling methods were completed. As a result of all these projects, successful results were obtained, as in the retail loan model.
Today, many models in Akbank are developed with machine learning technology and relevant teams are working to increase their usage area day by day. In order to ensure the continuity of the competence provided to Akbank by this project, many related teams received training at different levels on machine learning and we can say that we now have a very competent team in this field.
Finally, it has been on the agenda for many years that smart cooperation with different players in the innovation and entrepreneurship ecosystem can provide significant benefits to the large corporations, and as Akbank, we will continue to explore such collaborations in every field.
[1] https://www.mckinsey.com/industries/financial-services/our-insights/ai-bank-of-the-future-can-banks-meet-the-ai-challenge#:~:text=The%20potential%20for%20value%20creation,%2C%20annually%20(Exhibit%201)
