How does AI predict loan defaults?

Banks approved loans using a three-digit credit score for decades. The score captured some useful signals, but it missed far more than it measured. A small business owner with irregular income but zero late payments looked the same on paper as someone three months from default. A recent graduate with a thin file was indistinguishable from a genuine credit risk.

Predictive machine learning changed that calculus. Modern default prediction models process hundreds of variables at once and learn which combinations actually predict non-payment, not which combinations a committee decided looked risky in 1989. McKinsey's 2021 research found institutions using machine learning for credit risk reduced default rates by 20–30% versus traditional scoring alone. That is not a marginal improvement. For a mid-size lender with a $500 million portfolio, 25% fewer defaults can mean $15–$20 million in annual loss recovery.

The short answer is: more than most borrowers expect, and structured very differently from the way humans think about creditworthiness.

Traditional credit scoring draws on a narrow set of inputs: payment history, outstanding balances, length of credit history, new accounts opened, and credit mix. FICO, for example, reduces all of that to a 300–850 score using fixed weights that have not fundamentally changed since the late 1980s.

Machine learning models start from the same core inputs but extend well beyond them. A lender running a modern default model might feed it hundreds of variables across several categories.

Transaction behavior is one of the most predictive. How often does the borrower overdraft? What share of income goes to recurring obligations? Do they make minimum payments or pay balances down? A pattern of paying the exact minimum every month, for instance, predicts future distress better than the credit score itself.

Income stability matters more than income level. Two borrowers earning $80,000 per year look identical on a traditional form. But one has twelve regular direct deposits over the past year and the other has four lump sums with three-month gaps. The model sees them as very different credit profiles.

For business loans, the model typically ingests revenue trends, accounts receivable aging, payroll consistency, and supplier payment timing. A company that consistently pays suppliers 15 days late is flagging cash flow stress long before the credit report catches it.

Alternative data adds a further layer. Depending on jurisdiction and lender policy, models may also incorporate rental payment history, utility payment records, and in some markets, mobile phone top-up patterns. The World Bank's 2021 Financial Inclusion report found that alternative data meaningfully extended credit access in markets where 40–60% of the adult population lacked traditional credit history.

The output of a default prediction model is a probability: this borrower has a 7.3% chance of missing three consecutive payments within the next 18 months. Getting to that number involves a specific sequence of steps.

The model is trained on historical loan files where the outcome is already known. If the dataset contains 500,000 past loans, 40,000 of which defaulted, the model learns to find patterns in the 460,000 that did not default and the 40,000 that did.

Most production-grade default models today use gradient boosting algorithms, with XGBoost being the most widely deployed as of 2022. Gradient boosting builds a series of decision trees, each one correcting the errors of the last. The final prediction is an ensemble of hundreds of those trees voting together. A 2020 benchmark study published in the Journal of Risk and Financial Management found XGBoost outperformed logistic regression (the traditional statistical approach) by 12–18 percentage points on default detection accuracy across five different lending datasets.

The model assigns an importance weight to each variable based on how much it improves prediction accuracy. Payment-to-income ratio typically ranks near the top. So does the trend in credit utilization over the past six months. The exact weights are learned from data, not set by a committee, which is both the power of the approach and the source of its fairness risks.

Once trained, the model scores each new applicant in milliseconds. The score maps to a probability, and lenders set their own threshold. A bank might approve any applicant below 5% predicted default probability, flag 5–12% for manual review, and decline above 12%. Those thresholds are business decisions, not algorithmic ones.

For most lenders, yes, but the comparison requires some care.

Credit scores are designed to be broadly applicable. A FICO score works reasonably well across millions of borrowers with very different profiles because it uses the same formula for everyone. Machine learning models, by contrast, tend to be more accurate within a specific lending context but can degrade when applied outside the data they were trained on.

A model trained on consumer auto loans at a regional US bank will outperform FICO on that exact population. The same model applied to small business loans in a different market will likely underperform FICO because the training data does not match the new context.

The McKinsey Financial Services practice's 2022 report on credit risk noted that lenders combining machine learning scores with traditional credit scores outperformed either approach alone. The two methods capture different signals. The traditional score provides a stable, auditable baseline. The ML model catches the behavioral patterns the score misses.

For lenders considering the build-versus-buy question, the choice of implementation partner matters considerably. Building a production-grade default model requires data engineers, ML engineers, model validation specialists, and ongoing monitoring infrastructure. A global engineering team can deliver that infrastructure for $25,000–$40,000, where a US-based data science consultancy typically bills $90,000–$150,000 for equivalent scope.

This is where the technology gets genuinely complicated, and where many lenders have made expensive mistakes.

The core risk is straightforward: if historical lending data reflects discriminatory decisions (and in most markets, it does), a model trained on that data will learn to replicate those decisions. Race is not a permitted variable in US lending models under the Equal Credit Opportunity Act. But zip code, which correlates strongly with race due to decades of residential segregation, often is permitted. The model does not need to see race directly to produce racially disparate outcomes.

The US Consumer Financial Protection Bureau found in a 2022 supervisory review that several major lenders' automated underwriting systems produced denial rates for Black and Hispanic applicants that were 40–80% higher than for comparable white applicants, even after controlling for income and credit score.

Fairness in machine learning is not a single metric. It is a set of tradeoffs, and different stakeholders disagree on which one matters most.

Calibration asks whether the predicted default probability is accurate across demographic groups. If the model predicts 8% default for a group and 8% of that group actually defaults, the model is calibrated. But calibration can coexist with disparate approval rates.

Demographic parity asks whether approval rates are equal across groups. Achieving demographic parity often requires the model to apply different thresholds to different groups, which is itself legally contested.

Individual fairness asks whether two similarly situated borrowers receive similar decisions. This is intuitively appealing but practically hard to define and enforce.

For a non-technical founder building a fintech product, the practical implication is this: automated credit decisioning requires ongoing fairness auditing, not just model accuracy monitoring. A model that was fair at launch can drift as the underlying population shifts. The engineering cost of monitoring is smaller than it sounds: a well-built monitoring system can be set up for $8,000–$12,000 and run automatically without manual intervention. The regulatory cost of getting this wrong is not small at all.

Building responsible default prediction infrastructure is exactly the kind of problem a specialist data engineering team handles better than a generalist agency. The model logic and the fairness monitoring have to be co-designed, not bolted on afterward.

Book a free discovery call