The area of credit decision-making isn’t static. As new technologies develop, underwriters continue to create new ways to improve the critical components of this process, including predicting the likelihood of default, pricing loans, and efficiently managing clients throughout the life of a loan.
Presently, Generative AI (GenAI) transforms credit decision-making at an astonishing velocity, and its real-world impact is just beginning to manifest. The article in this issue, Romain Mazoue from the Group’s Chief Risk Officer, Younited, examines the development of credit decision-making. He also delves deep into Artificial Intelligence (AI) as well as GenAI applications that are transforming the sector, and gives an insight into the future of these technologies as they are integrated into the operations of lenders.
Table of Contents
Important lessons to take away
- The transition from manually-driven to automatic credit decision making: Over the course of time, credit decision-making processes have evolved from manual procedures dependent on human judgment towards highly automated, information-driven processes that are powered through AI or machine learning (ML).
- Making use of diverse data for better results: Thanks to AI and ML, as well as greater data access, lenders can analyse massive quantities of traditional and non-traditional data, like the activity of social media and utility payments, allowing for more precise and more personalized credit assessments.
- GenAI enhances decision-making capabilities: Generative AI creates a new layer of sophistication through the interpretation of unstructured data, providing insights and enabling human-like interactions, improving the capabilities in AI as well as ML in the process of credit decision-making.
- Resolving the challenges to long-term sustainability: Despite its potential massive adoption, GenAI will require financial institutions to reduce the risk of bias, governance, and reliability of the system to ensure ethical and fair applications.
- Future-proofing customer-centric lending by using: GenAI responsibly, lenders can simplify credit decision-making and foster the development of new ideas, inclusiveness, and efficiency. This will help redefine the standards of financial services that are geared towards customers. services.
A brief review of the history of credit decision-making
The process of credit underwriting has seen a significant change in the past, moving from subjective, manual evaluations towards highly computerized and information-driven systems. Below is a comprehensive timeline of the decision-making process for credit underwriting
1. The introduction of algorithms for credit scoring (1950s to the 1980s)
Manual Underwriting and expert judgement. In the beginning days of the world of underwriting credit, the decisions were taken manually by underwriters or credit managers. They would look over the applicant’s financial records, including earnings, debts, and personal characteristics, in order to judge their creditworthiness. This process was extremely subjective and was being influenced by biases, such as gender, race, or personal opinions about the character of an applicant. The manual method was slow and susceptible to a variety of inconsistencies.
The introduction of the credit scoring algorithms
The first significant change in the way credit was underwritten was the introduction of models for credit scoring in the 50s. FICO (Fair Isaac Corporation) was established in the year 1956, in 1956 by Bill Fair and Earl Isaac established an early credit scoring model that were designed for lenders. These models were initially designed specifically for specific companies and were not widely adaptable to different industries. Three other major players in the field of credit scoring were also developed during this time.
- Experian: The business was founded in the year 1980 as CCN Systems in Nottingham, England. Experian’s U.S. branch of Experian has its origins dating back to the year 1897, which was the time that Jim Chilton created the Merchants Credit Association.
- Equifax was founded in the beginning in 1899 as the Retail Credit Company 1899 by Cator and Guy Woolford in Atlanta, Georgia. They changed their name from Equifax in 1979.
- TransUnion was founded in 1968 as a holding company parent of the Union Tank Car Company. It was in 1969 that TransUnion bought the Credit Bureau of Cook County, which marked its entry into the world of credit reporting.
In the 1970s, the Fair Credit Reporting Act (FCRA) began to regulate the acquisition of personal credit information as well as the access of customers to credit information within the USA.
In the late 1980s, credit bureaus started to digitize data from consumers, and credit scoring models were more widely used. In 1989, FICO introduced its initial “universal” credit score model that could be utilized by any lender using credit reporting agencies like Equifax, Experian, or TransUnion. This was the start of a more uniform and objective approach to assessing the risk of credit.
At that moment, fixed pricing was commonplace despite huge differences in costs for customers (specifically, costs of risk) and APR on loans only changing according to expert judgement.
2. Advanced engineered credit decisions (1990 until the 2000s)
Wider use of credit scores: In the late 1990s, sophisticated underwriting methods began to emerge in the form of credit card products, but later expanded to other lending options, specifically in the mortgage sector. They were developed by US government-sponsored companies such as Fannie Mae and Freddie Mac. These systems were designed to speed up the underwriting process through automated evaluation of a borrower’s creditworthiness using established criteria, such as debt-to-income ratios and credit scores. The new underwriting system drastically reduced the time needed to approve loans and also improved the consistency of decision-making.
In the first decade of the 2000s, technological advances in computing and analytics for data led to the creation of more advanced statistical models, including logistic regression, which can forecast default risk more accurately. Banks started hiring statisticians and data scientists to build these models that were trained on huge data sets that contained hundreds of variables that relate to borrower behaviour. The credit decision engines were slowly integrated into the tech stacks of financial institutions, which allowed real-time decision-making on a broad range of data on borrowers.
Parallel to this, a variety of factors contributed to increasing the utilization of credit scoring, as well as an improvement in the pricing of loans:
1. US regulation and credit scoring options
The US Congress adopted the Fair and Accurate Credit Transaction Act in 2003. The law allowed customers to ask for and receive a free credit report every 12 months from any of the three national consumer credit reporting firms (Equifax, Experian, and TransUnion). However, adoption of this policy was slow because of the absence of awareness and marketing.
2. Basel II Regulation
The Basel II Internal Ratings-Based Approach (IRBA) requirements were announced at the end of June 2004. This framework permitted banks to utilize their own internal ratings to evaluate credit risk and determine the capital requirements of regulatory authorities, provided that they met a set of minimum requirements and had approval from their supervisors in their country.
Through allowing banks to design and implement their own models by allowing them to develop their own models, the IRBA encouraged institutions to enhance their risk management strategies. Banks were encouraged to implement advanced modeling techniques to help them better manage their exposure to credit risk by estimating risk parameters. These models enabled an improved assessment of risk to credit, which included both non-observable and visible risk variables.
The IRBA introduced strict requirements from the regulator to validate models and approve their use. Banks had to show that their models are solid, accurate, and in line with the regulations. This led to greater scrutiny of model practices and also encouraged continual improvement in the development of models and validation procedures.
3. Capital One and price based on risk
Early innovators went above credit scoring to perform complex modeling of a variety of other factors, including churn, the value of a customer’s lifetime, and profitability scoring.
In the early 1990s, Capital One developed an innovative marketing strategy that targeted customers who focused on customer profitability analysis. The result was incredible success as a major provider of credit cards. It maintained its position through investing in its infrastructure as well as staff and constantly enhancing its skills using a “test-and-learn” method. Capital One’s “information-based” strategy consisted of applying a pricing model based on risk.
3. Integration of AI and alternative data (2010s)
The next major leap in credit underwriting occurred with the introduction of Artificial Intelligence (AI) and Machine Learning (ML) technologies. These tools let lenders examine huge amounts of traditional and non-traditional information (e.g., social media activity or utility payments) for more precise credit decision-making.
AI-powered systems could evaluate the reliability of borrowers beyond traditional measures like credit scores, expanding credit options for people who have a limited credit history. AI can also improve the risk assessment process by identifying the patterns of behavior in borrowing that conventional models overlook. But, there are some challenges in the new model generation, particularly that there is no transparency, as well as the risk of biases in the model that could lead to discrimination.
The increasing use of data from other sources has led to the rise of AI models. By leveraging larger datasets of diverse and non-linearly correlating datasets, AI models have shown superiority to the traditional models of statistical analysis. To make credit-related decisions, alternatives typically comprise (i) open bank records, (ii) utilities and rental history of payments, (iii) open admin data, as well as (iv) the digital footprint. This strategy has proved especially useful for populations that aren’t served like new immigrants or workers in the gig economy, who would otherwise be disqualified from traditional lending programs.
4. Real-time decision-making and alternative information (mid-2010s until 2020s)
Today, many lenders employ real-time data analytics in order to make rapid decisions about loans. Automated systems are now able to integrate different data sources — like rent payment history or utility bills to assess the creditworthiness of borrowers who do not have a strong credit background. This change has led to lending becoming more inclusive, but also ensuring that there are strict procedures for managing risk.
Furthermore, modern decision-making systems permit greater flexibility in decision-making by allowing for configurable guidelines that are able to be modified without significant changes. SaaS systems for making decisions, as an example, let lenders design and validate highly segmented real-time decision workflows and models that are adapted to the needs of their customers.
5. GenAI’s emergence (2022)
Although AI models have been around for a long time–in the 1980s, with logistic regressions — they were restricted by the smaller data sets and lower computational power. The bases that underpin GenAI models are relatively new in their development, starting with the creation of Transformer models by Google in 2017 and the launch of BERT in. BERT was a major technological breakthrough in the field of processing languages and revolutionized the capability to recognize context using its transformer-based structure
On November 20, 2022, OpenAI launched ChatGPT, which quickly became a groundbreaking product thanks to its ability to create text that resembles human language. In just the span of two years, ChatGPT grew to more than 100 million users, making it the fastest-growing user application ever. A couple of years on, GenAI models have continued to improve, with a variety of new major players competing against ChatGPT: Gemini, Android, and Mistral, for example.
In the last two years, the effectiveness of LLM models has been improving at an impressive rate, with these notable improvements:
- Models with multimodal capabilities have widened their functions beyond text generation to include code generation and image generation, and have broadened their use.
- Model customization and optimization of performance through RAG, fine-tuning along with prompt engineering, has grown in maturity.
- There is now a greater integration of GenAI agents into the majority of software applications via chatbots, Copilot, and more.
- The efficacy and capacity of AI models that are generative AI models have been improved, and they are now able to handle greater workloads and offer quicker responses, which allows for real-time interactions.
- Smaller versions or open-source versions of GenAI models offer better portability and efficiency in cost.
Furthermore, hardware capabilities have been a part of the Gen AI hype, with Nvidia becoming the most valuable global corporate valuation, ranging between $500-$3,200 million!
These examples show how quickly GenAI has advanced from being a new technology to an incredibly popular tool in various industries, improving effectiveness and personalization across a variety of applications.
In the last several years, structural trends have changed the way decisions about credit are made in three stages. We began by shifting from deterministic behaviour using expert advice to probabilistic behaviour using deep learning and machine learning models that can learn from data, without being explicitly programmed. The third stage is GenAI models, which utilize content to generate new content as well as process complicated workflows. The three models are able to be utilized in a complementary way.
Another trend that is noticeable is the shift from proprietary software, such as SPSS and SAS, to open-source software and libraries (R, Python). Open-source communities have encouraged the development of new technologies and have also aided in providing new algorithms and tools, such as XG Boost or CAR, accessible.
AI GenAI and AI GenAI employ cases for credit decision-making
Focusing on the third and second steps mentioned above with respect to credit decisions and loan applications, AI is a great option. Generative AI (GenAI) offers distinct capabilities that complement one in order to improve the process of loan approval.
AI capabilities in credit decision-making
Traditional AI models operate as deductive minds to make decisions. They can be utilized in various steps of the credit decision-making framework, such as:
- predictive analytics, and risk assessments: AI leverages predictive analytics to evaluate creditworthiness through analyzing huge amounts of data, such as the patterns of spending and financial behavior that are based on open banking. This allows lenders to make quick and informed decisions, reducing the likelihood of defaults. They can spot nonlinear connections and complex interactions in an enormous amount of data that logistic regression may overlook.
- The AI system for fraud detection excels at detecting abnormal patterns or anomalies in real-time and helps to stop fraudulent activities before they affect financial institutions.
- Efficiency and automation: AI models can help automate repetitive tasks, such as compliance and data entry, improving processing time and decreasing human errors. This results in faster approval of loans and improves the efficiency of operations. In addition, AI models can automate the development of models through automated feature selection as well as model selection, which reduces the requirement to use manual intervention.
GenAI capabilities in credit decision-making
GenAI models are similar to the inductive brain because they are able to demonstrate innovative thinking and problem-solving skills. The new capabilities that GenAI models provide can be summarized in the following manner:
- Learn the meaning of a process and reverse engineering. solving problems.
- Converse: Human-like dialogue, human-like language interactions, as well as instruction.
- Create: Images, writing codes, the generation of multimodal content.
- Coach: Real-time coaching.
- Command: Determination of the best workflows.
While this brand’s capabilities are not specific, however, we are seeing increasing instances of positive Gen AI implementations to tackle the issues of risk management:
- Development of products: GenAI-enabled programming development, code testing, documentation generation, and identifying patterns that will help in the development of new products.
- Distribution and sales customer segmentation: Analysis of sentiment, hyper-personalized content, offers, and optimization of pricing.
- Operations: The following are the main ones: Onboarding, credit score collection, servicing, along enhanced underwriting.
- Risk and compliance Monitoring of risk and compliance: Early warnings, suspicious activity reports creation, monitoring contracts, prevention of fraud and money laundering, and detection document summaries for credit underwriting, proactive recovery strategies, as well as the management of compliance and risk. All of these fall in the scope of this post.
- Support function: Employee self-service, Knowledge smart search, as well as document creation.
- Document handling: Handling unstructured information, like customer feedback, or more complex financial documents, through a summary of data, extracting the most important details, and automating the generation of documents. This technology improves the process of credit decision-making by making it easier to make decisions without manual effort and reducing errors.
- Customer interaction and personalisation: Enhancing customer service via personal interactions with chatbots via chatbots providing 24/7 support and specific suggestions. GenAI also creates customized repayment plans that are based on the borrower’s financial background.
AI versus GenAI
There are a variety of areas in which both kinds of AI can complement one another, like the recommendation engine and forecasting of demand. In addition, GenAI will enhance the traditional AI opportunities, but not substitute them!
In short In summary, whereas traditional AI is focused on predictive analytics and fraud prevention, and automation to improve the process of making credit decisions, GenAI offers advanced capabilities for handling unstructured data, improving personalization, and creating artificial data to train models.
These technologies complement each other to enable more efficient, accurate, and more customer-focused lending practices.
What’s next for credit decision-making after the day of GenAI?
Model risk
Model risk could be the negative side of the increased sophistication and complexity of models for credit decision-making that incorporate increasing parameters. Challenges such as lack of transparency, the risk of discriminatory bias, sensitivity to data quality, and limited explainability become more pronounced, posing significant concerns–particularly in the highly regulated financial industry.
AI Act and other regulations
Europe’s Artificial Intelligence Act (AIA) has an important impact on credit models for making decisions because it classifies AI systems used in credit scoring and creditworthiness assessment in the category of “high-risk.” The AIA requires strict standards for high-risk models, such as strong risk management with transparency, precision, human oversight, and preventive measures to avoid discrimination and unfair outcomes. These guidelines aim to limit possible risks such as bias or ambiguity, as well as ensure that credit models powered by AI work with integrity.
Through promoting accountability and fairness, AIA is seeking to ensure that the advantages of AI, like increased efficiency and inclusivity, as well as the protection of fundamental rights and the trust of consumers in financial decision-making. However, this could create challenges for European model designers because compliance issues and costs can hinder innovation and performance.
More reliable GenAI models
GenAI models have been known to occasionally deliver unexpected results. In highly sensitive and regulated processes like decisions on credit risk, this issue could pose a challenge. Technologies like Retrieval Augmented Generation (RAG) could solve problems like hallucinations and make sure that the AI-generated outputs are rooted in facts. In addition, GenAI’s capability to give clear explanations to decisions can increase confidence among regulators and consumers and provide applicants who have been rejected with more empathetic feedback.
Increased adoption and impact on the credit decision-making process
As open finance continues to grow, GenAI will play a crucial role in integrating data from multiple financial institutions in order to give comprehensive credit scores. With increased reliability and gradual changes to internal workflows as well as software to support the GenAI-based model, we are expecting major changes in the operational process. GenAI agents and models will become more independent in the credit decision-making process, with the potential of safety nets that are managed dynamically, activated by human-in-the-loop checking or controlling.
GenAI-driven credit decision-making may result in more precise and more inclusive decision-making in lending, increasing the availability of credit to people who are not served and lessening the risk of default. But the widespread acceptance of GenAI will need financial institutions to deal with issues related to bias, as well as governance and the system’s reliability.GenAI models cannot just streamline credit decision-making, but also encourage creativity, inclusion, and efficiency in the financial industry and set new standards for lending that are centered on customer practices.
Disclaimer
The information contained in this article does not constitute, and is not intended to provide, any professional advice. Instead, all information, data, and materials are intended to be used for general informational and educational purposes only. So, prior to making any decisions based on this information, we suggest that you speak with appropriate experts.