Fintech
Why AI Lending Models Need More Than Accurate Credit Scores
Add Securities.io to your preferred sources on Google
From Arbitrary Decisions To Credit Scores
When lending money, banks and other financial institutions always had to worry about borrowers not being able to pay it back. This is why many loans are backed by assets like real estate or business assets that the bank can seize if loan repayments are not made. Still, the decision was in the hands of bank personnel, who decided whether to grant the loan.
Over time, the individual decision of a bank employee was replaced by increasingly sophisticated filters based on internal risk models and mathematical formulas, making it more standardized and less arbitrary.
“Part of the job of the credit manager was to try to make inferences about the person’s personality, whether they seem to be organized and mature and responsible. A credit manager could reject an applicant if they suspected they were an alcoholic, or for more blatantly discriminatory reasons related to someone’s race, sex, or, for women borrowers, their marital status.
Josh Lauer – Professor at the University of New Hampshire and the author of the book “Creditworthy.”
However, it took until 1958 for FICO, founded by Engineer Bill Fair and mathematician Earl Isaac, to introduce the modern standardized credit score. And in 1989, the company launched the first general-purpose, nationwide credit score, ranging from 300 to 850, which quickly became the US national lending standard.
The Limits Of Credit Score
A modern credit score is based on five distinct data categories:
- Payment history(35%).
- Amount owned (30%).
- Length of credit history (15%).
- New credit and inquiries (10%).
- Credit mix (10%).
This method made the estimation of creditworthiness a lot more objective and fair than the previous method. But of course, it has its limitations.
For example, it does not include any information about an individual’s behavior outside of things related to past and present debt. It also does not consider information like place of living, income, sector of employment, assets owned, etc.
A modern credit score is also built around the assumption that banks give loans to individuals. But modern peer-to-peer (P2P) loan systems and other FinTech financial services function differently.
In parallel, progress in AI has made it possible to integrate complex, disparate data into a cohesive decision model. So it is no surprise that AI is making quick progress in improving models that predict loan defaults for a new generation of digital lenders.
A new study by two researchers at the University of Agder in Norway examined credit risk management across peer-to-peer lending platforms. They found that research has largely divided into two camps: one focused on improving default prediction through machine learning, and another focused on understanding why borrowers default. Both approaches, however, remain constrained by aging datasets, inconsistent definitions, and limited geographic coverage.
The researchers published their findings in Technological Forecast, Social Change1, under the title “Credit risk management in peer-to-peer lending platforms: A systematic literature review”.
The Rise And Challenges Of P2P Lending
Quick Growth, New Risks
The core concept of P2P lending encompasses disintermediation, where platforms function as marketplaces rather than risk-bearing entities, facilitating direct matching of supply and demand for funds. Many new FinTech companies are leading the charge in developing these markets, as well as P2P lending protocols using blockchain to bypass the need for a centralized lending provider entirely, alongside other DeFi (decentralized finance) technologies.
P2P lending is estimated at a $215B market size in 2026, and expected to grow at a 21.9% CAGR until 2033, with the majority of lending provided by individual investors.
Source: Coherent Marketing Insights
This is, however, a very new sector and is still poorly regulated, with risks ranging from poor estimation of the real default risks to outright fraud.
Analyzing P2P Literature
The researchers identified a total of 433 studies relevant for their analysis about P2P lending and credit risk management, of which 280 were analyzed further in-depth.
The scientific studies on the topic started to really pick up in 2014 and have reached a plateau in 2020. Geographically, most researchers in the field originate from China, the USA, India, and Indonesia, with China having published 39% of total publications.
Most of the ongoing research is focused on P2P lending in general, credit risks, and AI-related technologies like machine learning. Many of these studies analyze the roles of both soft information (social networks) and hard information (interest rates) in predicting defaults.
The most common paper topic is assessing predictive accuracy by applying various machine learning models, accompanied by data, variable, and model-level considerations.
The second most common topic is identifying the determinants of default likelihood and default severity.
Other studies propose new systems of credit risk management or discuss what remains under-researched.
P2P Lending And AI
Finding New Risk Factors
Because P2P lending does not rely solely on traditional credit scores, machine learning and, more recently, other AI technology are used to identify default risks more accurately.
For example, fuzzy cognitive maps building on soft computing approaches and neural network theory were used to create a model for credit risk assessment.
Regardless of specific method, studies often show that machine learning outperforms traditional credit scoring methods in terms of accuracy.
Overall, it appears that individually, borrowers’ credit scores, education, marital status, and job titles are typical indications of borrowers’ creditworthiness.
Another factor, based on “social capital theory”, can help as well: membership in a given social network can be a determining factor in identifying individuals less likely to default on their loans.
The researchers found a U-shaped contribution by age in one study, showing that middle-aged applicants (above 45) have higher default risk because of their heavier economic burden from supporting both their children and retired parents.
They also found that female borrowers are charged higher interest rates, despite their default rates being significantly lower than those of male borrowers.
Unsurprisingly, longer employment history, higher income, and lower debt ratio are favored.
Property ownership, including houses and cars, also boosts borrowers’ repayment ability, while rentals and outstanding car loans weaken it.
They also found that “soft” metrics can be predictive, albeit not necessarily actionable by P2P lending platforms.
“One study tried tapping into the character of borrowers, showing that default was more likely among students who were delayed in returning books to a university library and failed a higher proportion of the courses they took.”
In total, the study found that the most important factors to take into account are:
- At the borrower level:
- Demographic.
- Financial performance.
- Solvency.
- Credit score.
- Social capital.
- Psychological indicators.
- Loan level (amount, repayment period, purpose, authentication, and description length).
- At the lender level:
- Demographic indicators.
- Herding.
- At the platform level:
- Special collection actions.
- Price regimes.
- At the macro level:
- Regulation.
- Economic development.
- Employment.
- National culture.
- Crisis periods.
Black Box Evaluation
A recurring issue with AI is that neural networks and machine learning tend to create so-called “black box” systems, where the reasons for a given result are far from clear even to the developers of the system, and even less to later operators.
This can be a serious issue for P2P lending using AI.
First, this creates serious dangers of default risk miscalculation, with no way to identify the problem on time.
Second, this can pose serious regulatory risks, as many of an individual’s characteristics like race or gender are generally considered discriminatory and illegal to integrate into the decision process of granting a loan or not.
Building An AI Credit Score
Building an AI credit score requires using the right dataset for the AI training.
The researchers argue that most AI evaluation of creditworthiness should integrate multiple time frames. Short time frames (<30 days) help identify early warning signs, while long time frames (120 days or more) give a better picture of the overall result of the lending.
Recovery, or the probability of repaying overdue sums upon the occurrence of default, is also important data.
Lastly, a perfect AI system should also continuously adjust for changing economic conditions and demonstrate that its underwriting remains reliable outside the historical datasets on which it was trained.
Overall, the study points toward the direction of the “perfect” AI-driven credit score to integrate as many relevant variables as possible, including loan structure, platform policies, investor behavior, employment conditions, regulation, and macroeconomic shocks. However, making such a decision process transparent and fair might be a significant challenge with many of the current AI technologies and training procedures.
Investing In AI Credit Score
Upstart Holdings, Inc
As lending and credit scores evolve, new companies are emerging to provide better alternatives to traditional credit scores and lending practices.
One of them is Upstart (UPST ), a lending marketplace powered by AI, launched in 2012, much before AI became a center of tech conversation.
UPST Price Chart
The idea behind Upstart was that the existing credit score system is inefficient and outdated. With a lot more data available, it had to be possible to identify loan risks better and, as a result, provide cheaper loans to a large portion of the population.
“Our personal loan model is meaningfully more accurate than a traditional credit model at evaluating borrower risk. We believe years of modeling improvements lie ahead.”
This early start gave the company an abundant source of data on which to train its own AI and leverage the recent progress made in AI training methods.
Source: Upstart
This means Upstart’s method can identify people with high FICO scores but, in practice, have a high risk of defaulting on their loans. And conversely, people with low FICO scores are not that likely to default.
So with the use of AI and data, Upstart can achieve outstanding results for a given cohort. For example, it can either manage to offer more loans for the same level of risk, rising from a 27% approval rate to 67%. Or it can reduce the losses significantly, from a 9% to 5% default rate, or almost half as much.
Source: Upstart
In addition, using an algorithm instead of a loan officer means the process is automated. This is both less expensive to do and much quicker and pleasant to apply for a loan.
Upstart’s marketplace connects borrowers with banks, credit unions, and institutional capital. It reported that institutional investors purchased 64% of the principal originated through its marketplace in 2025.
Upstart serves an absolutely massive market, with $25T in credit originated each year globally. Of this sum, up to $1T in revenue is available to the platforms that originate and service that credit.
In 2025, the company originated 1.5 million loans (up 115% year-to-year), for a total of $11B (up 86% year-to-year), and total revenues for the company of $1B (up 64% year-to-year). The company is also growing quickly in new segments, with, for example, auto loan origination growing 5x (from $45M in Q4 2024 to $263M in Q1 2026), and home loan origination growing 6x (from $27M in Q4 2024 to $143M in Q1 2026).
So even if large, these numbers pale in comparison to the total potential of advanced credit score methods and AI-driven loan assessment.
Source: Upstart
The recent growth numbers and having been built as an AI-first company are strong positives for the future of Upstart, even if the stock price has not really rewarded this success, with prices at the same level as in 2022, and much below its 2021 peak during the pandemic.
In part, this is because there are also a few risks to Upstart’s business model.
The first is that its model can deteriorate, as economic conditions move beyond training data and its AI could miss new trends.
Another issue is increasing regulatory pressure for explainability, fairness, and demographic performance reporting, as even if inadvertently, AI could bring back discriminatory practices previously banned from the financial industry.
Lastly, the company still remains dependent on institutional funding remaining available across credit cycles, with signs of stress in the financial system building in the context of high debt levels and a tense geopolitical situation.
Still, Upstart is ideally placed to become a key AI lending marketplace and underwriting infrastructure provider, a necessary evolution of the initial FinTech and P2P networks’ push for new lending practices.
And investors might want to get exposure to a company that has been radically changing how loan underwriting is performed early on.
Latest Upstart Holdings (UPST) Stock News and Developments
Study Referenced
1. Fengya Zhu and Rotem Shneor. Credit risk management in peer-to-peer lending platforms: A systematic literature review. Technological Forecast, Social Change. November 2026. Article: 124831. Volume 232. 10.1016/j.techfore.2026.124831
