Three approaches to deciding who gets a loan, and why better algorithms mean better access
Every time a bank approves a loan, it is making a bet: that you will pay the money back, plus interest, on schedule. If the bank is right, it earns a return. If it is wrong, it loses the money it lent you. The entire lending industry is built on making this bet correctly, millions of times over.
Banks borrow money cheaply (from depositors and wholesale markets) and lend it at higher rates. The difference, called the spread, is their primary revenue. But some borrowers do not pay back. When that happens, the bank loses not just the interest but the principal. Too many defaults and the bank is insolvent.
Lenders charge different rates to different people, and the reason is math. A borrower who is very likely to repay gets a low rate (say 5%). A riskier borrower gets a higher rate (say 15%). A borrower the lender expects will default gets denied entirely.
The lender's goal is not to avoid risk entirely. A lender that only approves the safest 10% of applicants leaves enormous revenue on the table. The goal is to price risk accurately: set rates that compensate for actual default probability and lend to as many people as possible at rates that work for both sides.
How do you evaluate millions of loan applicants quickly? For the past 35 years, the answer has been the credit score. The most common is the FICO score: a number between 300 and 850 summarizing your credit history. High score means you have a track record of paying back debt on time. Low score means you do not.
Lenders set cutoffs. Above 720 is usually "prime" (best rates). Between 620 and 720 is "near-prime." Below 620 is "subprime" (high rates or denial). These cutoffs determine who gets access to affordable credit and who does not.
Credit scores work well for people with long, documented credit histories. They fail for everyone else.
According to the Consumer Financial Protection Bureau, 26 million American adults are "credit invisible" (no credit file at all) and another 19 million are "unscorable" (too thin to generate a score). These 45 million people are not necessarily bad risks. Many are recent graduates, immigrants, cash-economy workers, or people who have simply never borrowed. They are invisible to the system.
In the years before the 2008 financial crisis, mortgage lenders approved millions of borrowers who could not actually afford their loans. Many had credit scores that looked adequate: they had some credit history and were current on existing accounts. But the scores did not capture their ability to pay under stress: what happens when the adjustable rate resets, when housing prices drop, when they lose a job.
When housing prices fell, millions defaulted. Many were recent immigrants and first-time buyers with "thin file" credit profiles who had been sold mortgages they did not understand and could not sustain. The scoring system failed both the borrowers (who lost their homes) and the lenders (who lost billions).
The lesson: a score that does not capture real ability to pay is dangerous for everyone. The three approaches that follow represent three different attempts to score more accurately.
The FICO score was introduced in 1989 by Fair Isaac Corporation. Before FICO, lending decisions were local and subjective: a loan officer would interview you, check your references, and make a judgment call. This system was slow, expensive, and rife with discrimination. FICO replaced subjective judgment with a standardized number.
Select an applicant to see how FICO evaluates their credit profile.
Upstart was founded in 2012 by former Google employees with a hypothesis: machine learning models trained on thousands of variables could predict loan defaults more accurately than FICO's formula. They were right. The results earned a no-action letter from the Consumer Financial Protection Bureau in 2017, giving regulatory clearance to use education and employment data in lending decisions.
Upstart starts with the same credit data FICO uses, then layers on:
The ML model (gradient-boosted trees, not a single formula) processes over 1,600 variables versus FICO's ~30. It finds non-obvious patterns: correlations between specific education/employment combinations and default rates that no human analyst would think to code as rules.
Select an applicant to see how the two systems evaluate the same person.
ML models find any pattern that predicts default, including patterns correlated with race, even when race is not an input variable. ZIP code, university attended, and field of study can serve as proxies for protected characteristics. If graduates of historically Black colleges default at different rates (for any reason, including systemic inequity), the model will pick up that pattern.
The debate is ongoing: does ML lending reduce bias (by finding creditworthy people FICO misses) or amplify it (by finding new proxies for discrimination)? Evidence suggests it does both, depending on implementation. The regulatory framework is still evolving.
FICO requires credit history. Upstart requires credit history plus more. Tala requires neither. Founded in 2014, Tala operates in Kenya, the Philippines, Mexico, and India, serving people who have zero formal credit history. Its approach: use smartphone behavioral data to assess creditworthiness from scratch.
According to the World Bank, approximately 1.4 billion adults worldwide have no bank account. In sub-Saharan Africa, only about 55% of adults have any formal financial account. These people have legitimate credit needs (business inventory, medical emergencies, school fees) but no way to prove creditworthiness under traditional systems.
When a user applies for a Tala loan, the app (with permission) analyzes behavioral signals from their smartphone:
Select a borrower to see how phone behavioral data becomes a credit assessment.
Tala has disbursed over $4 billion in loans since 2014. Average loan sizes are small ($50 to $200) because these are first-time borrowers building a track record. Default rates are competitive with traditional microlenders. For millions of users, a Tala loan is their first interaction with formal credit.