Fintech Sector

Predicting Bank Loan Defaults for Profit Maximization

Predictive Modeling

Developing an AI model for loan default prediction and using it to increase bank profits?
Challenge Accepted!

Predicting Bank Loan Defaults for Profit Maximization

Process & Story

A company from the fintech industry was trying to develop an AI model that could predict bank loan defaults in order to use the information for profit maximization. The company has access to a large dataset of customers’ credit records and other related data on their customers’ activity that could potentially be used to automate credit scoring and streamline the decision-making process for credit approval.

Problem

In the current credit scoring systems, it is a very time-consuming process for agents to manually look through all of the documents and other information to decide on the loan approval. In addition, there was no comprehensive way to determine which customers would default on their loans without manually going through every case file. On the other hand, banks gain additional interest rates from clients late with their installments, so the credit decision should also be optimized for bank profit maximization.

Solution

To predict loan defaults, we applied logistic regression, which is a parametric machine learning algorithm that models the probability of falling into either one of the binary classes (client will default with his loan or have a clear payment status).

The model was trained with historical data on loan defaulted clients to learn the patterns which are related to defaults.

The results of the logistic regression showed that besides some apparent features like credit score, debt to income ratio, or duration of the credit history, there were also some other less obvious ones that proved to be significant for predicting the bank loan default (e.g., demographic data, homeownership).

Due to the fact that the existing dataset is not balanced, which means that there are many more customers with clear loan status than customers who default, we used the sampling method to address this issue. The sampling method is a special case of statistical inference where observations are selected from a population to answer a question about the whole population.

Another important part of the solution was to make it interpretable and easy to understand for business users, so they can easily identify which features have the highest impact on predicting defaults. If the model is too complex and difficult to comprehend, it will be hard for bank agents to use it as a useful tool for decision-making.

Case Study Schema Predicting Bank Loan Defaults for Profit Maximization

Tools

Python

Python

Tensorflow

Tensorflow

Google Cloud Platform

Google Cloud Platform

Have a similar project in mind?

LET'S TALK!

Challenges

Model explainability
Like many other fintech AI-based solutions, this model was required by law and the regulatory authorities to be explainable. Our goal was to develop a model that could step by step explain the results for all processed cases marking which input parameter had what impact on the likelihood of the case falling into one of the binary categories (loans settled on time; defaulted loans). We decided to use the logistic regression model for binary classification since it is highly interpretable by nature.

Model interpretability and parametrization
As the model was developed to be used for profit maximization, it was necessary to have interpretable results. This means that each parameter of our regression model should give us insight into the borrowers' payment behavior on time or not on time. It was also designed to provide the financial institution the freedom to alter the settings and feature importance to match broader company strategic decisions (e.g., how changing interest rate from current sales campaign affects the default risk assessment, loan performance, and overall company annual income). We provided a fully transparent solution with attribute level weight estimations towards default probabilities.

Training datasets imbalance
The overall number of people defaulting with their loans is relatively much smaller than the number of loans paid on time. This has consequences on the training process of the logistic regression model. The negative cases are much less represented than the positive ones, making it harder to train a good model that can predict defaults accurately, limiting the number of false-positive decisions. For this reason, we used various sampling methods to even out these numbers and reach an expected accuracy level for default prediction.

Fast framework and infrastructure
The model had to work with a large dataset, and predictions should be quick enough for banks to assess credit risk, provide accurate credit scores and make AI-backed decisions on their loans in minutes after receiving each new incoming loan application.

Our other projects:

Insurance Sector

Auto Insurance Fraud Claim Detection

Developing an AI model that can detect car insurance fraud based on car accident claims?
Challenge Accepted!

Fraud Detection Predictive Modeling

Fintech Sector

Loan Application Fraud Detection

Developing a model that can identify fraudulent loan applications for a financial institution?
Challenge Accepted!

Fraud Detection Predictive Modeling

Social Media Platform

Topic extraction and personalization models for idea exchange system

Developing an AI-based system that is able to extract and classify interesting ideas from articles and books and recommend them to social network users?
Challenge Accepted!

Natural Language Processing Recommender System

See more projects

Find us on

Need help with implementing AI in your business?

Let's talk blue circle

This site uses cookies for analytical purposes.

Accept Privacy Policy

In the interests of your safety and to implement the principle of lawful, reliable and transparent processing of your personal data when using our services, we developed this document called the Privacy Policy. This document regulates the processing and protection of Users’ personal data in connection with their use of the Website and has been prepared by Nexocode.

To ensure the protection of Users' personal data, Nexocode applies appropriate organizational and technical solutions to prevent privacy breaches. Nexocode implements measures to ensure security at the level which ensures compliance with applicable Polish and European laws such as:

  1. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation) (published in the Official Journal of the European Union L 119, p 1); Act of 10 May 2018 on personal data protection (published in the Journal of Laws of 2018, item 1000);
  2. Act of 18 July 2002 on providing services by electronic means;
  3. Telecommunications Law of 16 July 2004.

The Website is secured by the SSL protocol, which provides secure data transmission on the Internet.

1. Definitions

  1. User – a person that uses the Website, i.e. a natural person with full legal capacity, a legal person, or an organizational unit which is not a legal person to which specific provisions grant legal capacity.
  2. Nexocode – NEXOCODE sp. z o.o. with its registered office in Kraków, ul. Wadowicka 7, 30-347 Kraków, entered into the Register of Entrepreneurs of the National Court Register kept by the District Court for Kraków-Śródmieście in Kraków, 11th Commercial Department of the National Court Register, under the KRS number: 0000686992, NIP: 6762533324.
  3. Website – website run by Nexocode, at the URL: nexocode.com whose content is available to authorized persons.
  4. Cookies – small files saved by the server on the User's computer, which the server can read when when the website is accessed from the computer.
  5. SSL protocol – a special standard for transmitting data on the Internet which unlike ordinary methods of data transmission encrypts data transmission.
  6. System log – the information that the User's computer transmits to the server which may contain various data (e.g. the user’s IP number), allowing to determine the approximate location where the connection came from.
  7. IP address – individual number which is usually assigned to every computer connected to the Internet. The IP number can be permanently associated with the computer (static) or assigned to a given connection (dynamic).
  8. GDPR – Regulation 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of individuals regarding the processing of personal data and onthe free transmission of such data, repealing Directive 95/46 / EC (General Data Protection Regulation).
  9. Personal data – information about an identified or identifiable natural person ("data subject"). An identifiable natural person is a person who can be directly or indirectly identified, in particular on the basis of identifiers such as name, identification number, location data, online identifiers or one or more specific factors determining the physical, physiological, genetic, mental, economic, cultural or social identity of a natural person.
  10. Processing – any operations performed on personal data, such as collecting, recording, storing, developing, modifying, sharing, and deleting, especially when performed in IT systems.

2. Cookies

The Website is secured by the SSL protocol, which provides secure data transmission on the Internet. The Website, in accordance with art. 173 of the Telecommunications Act of 16 July 2004 of the Republic of Poland, uses Cookies, i.e. data, in particular text files, stored on the User's end device.
Cookies are used to:

  1. improve user experience and facilitate navigation on the site;
  2. help to identify returning Users who access the website using the device on which Cookies were saved;
  3. creating statistics which help to understand how the Users use websites, which allows to improve their structure and content;
  4. adjusting the content of the Website pages to specific User’s preferences and optimizing the websites website experience to the each User's individual needs.

Cookies usually contain the name of the website from which they originate, their storage time on the end device and a unique number. On our Website, we use the following types of Cookies:

  • "Session" – cookie files stored on the User's end device until the Uses logs out, leaves the website or turns off the web browser;
  • "Persistent" – cookie files stored on the User's end device for the time specified in the Cookie file parameters or until they are deleted by the User;
  • "Performance" – cookies used specifically for gathering data on how visitors use a website to measure the performance of a website;
  • "Strictly necessary" – essential for browsing the website and using its features, such as accessing secure areas of the site;
  • "Functional" – cookies enabling remembering the settings selected by the User and personalizing the User interface;
  • "First-party" – cookies stored by the Website;
  • "Third-party" – cookies derived from a website other than the Website;
  • "Facebook cookies" – You should read Facebook cookies policy: https://www.facebook.com/policy/cookies
  • "Other Google cookies" – Refer to Google cookie policy: www.google.com/policies/technologies/types/

3. How System Logs work on the Website

User's activity on the Website, including the User’s Personal Data, is recorded in System Logs. The information collected in the Logs is processed primarily for purposes related to the provision of services, i.e. for the purposes of:

  • analytics – to improve the quality of services provided by us as part of the Website and adapt its functionalities to the needs of the Users. The legal basis for processing in this case is the legitimate interest of Nexocode consisting in analyzing Users' activities and their preferences;
  • fraud detection, identification and countering threats to stability and correct operation of the Website.

4. Cookie mechanism on the Website

Our site uses basic cookies that facilitate the use of its resources. Cookies contain useful information and are stored on the User's computer – our server can read them when connecting to this computer again. Most web browsers allow cookies to be stored on the User's end device by default. Each User can change their Cookie settings in the web browser settings menu: Google ChromeOpen the menu (click the three-dot icon in the upper right corner), Settings > Advanced. In the "Privacy and security" section, click the Content Settings button. In the "Cookies and site date" section you can change the following Cookie settings:

  • Deleting cookies,
  • Blocking cookies by default,
  • Default permission for cookies,
  • Saving Cookies and website data by default and clearing them when the browser is closed,
  • Specifying exceptions for Cookies for specific websites or domains

Internet Explorer 6.0 and 7.0
From the browser menu (upper right corner): Tools > Internet Options > Privacy, click the Sites button. Use the slider to set the desired level, confirm the change with the OK button.

Mozilla Firefox
browser menu: Tools > Options > Privacy and security. Activate the “Custom” field. From there, you can check a relevant field to decide whether or not to accept cookies.

Opera
Open the browser’s settings menu: Go to the Advanced section > Site Settings > Cookies and site data. From there, adjust the setting: Allow sites to save and read cookie data

Safari
In the Safari drop-down menu, select Preferences and click the Security icon.From there, select the desired security level in the "Accept cookies" area.

Disabling Cookies in your browser does not deprive you of access to the resources of the Website. Web browsers, by default, allow storing Cookies on the User's end device. Website Users can freely adjust cookie settings. The web browser allows you to delete cookies. It is also possible to automatically block cookies. Detailed information on this subject is provided in the help or documentation of the specific web browser used by the User. The User can decide not to receive Cookies by changing browser settings. However, disabling Cookies necessary for authentication, security or remembering User preferences may impact user experience, or even make the Website unusable.

5. Additional information

External links may be placed on the Website enabling Users to directly reach other website. Also, while using the Website, cookies may also be placed on the User’s device from other entities, in particular from third parties such as Google, in order to enable the use the functionalities of the Website integrated with these third parties. Each of such providers sets out the rules for the use of cookies in their privacy policy, so for security reasons we recommend that you read the privacy policy document before using these pages. We reserve the right to change this privacy policy at any time by publishing an updated version on our Website. After making the change, the privacy policy will be published on the page with a new date. For more information on the conditions of providing services, in particular the rules of using the Website, contracting, as well as the conditions of accessing content and using the Website, please refer to the the Website’s Terms and Conditions.

Nexocode Team