What Actually Happens When a Borrower Stops Paying
Default risk is just the chance that someone who owes you money will not pay you back. That is it. It sounds like a very simple concept because it is one. The complicated part is measuring it accurately when the stakes are high. I spent years building credit models for institutional lenders. You learn pretty quickly that "simple" and "accurate" rarely live in the same spreadsheet. When you lend money, there are three separate failure modes you have to account for. The borrower might not pay on time. They might pay partially over a long stretch. Or they might disappear entirely. Each one feeds into the overall probability of default, but they behave very differently in practice. A student with a missed payment is not the same risk profile as a factory that stopped paying suppliers six months ago. These are two distinct situations, and the model has to treat them that way or you lose money.
The Story Of Default Risk A Simple Explanation For Everyone
The standard way to measure this is through Expected Loss. Expected Loss equals the probability of default multiplied by the exposure at default multiplied by the loss given default. PD is how likely they are to stop paying. EAD is how much you are owed when they do. LGD is how much of that you actually recover. If someone defaults on a secured mortgage, the house gets sold and you might recover most of it. If it is an unsecured credit card loan, you are probably looking at fifty to sixty percent loss on what they owe you. These numbers are not theoretical. I have seen lenders miss this distinction and underprice risk by millions. The formula itself is clean. The data underneath it is not. In my experience working on actual portfolios, the hardest part is getting reliable historical default data for the specific type of borrower you are looking at. If you are pricing corporate debt for small manufacturing firms in the Midwest, you are going to have a very thin dataset. Thin datasets mean the model will give you answers that look precise but are basically noise. I dealt with this directly when building a model for regional commercial real estate loans. The default rates in the training data were based on national averages from the early two thousands, which included the subprime crisis. Using those numbers for current pricing would have wildly understated risk because the portfolio composition had shifted entirely. The workaround was to layer in forward-looking adjustments based on local vacancy rates and refinancing walls. It was tedious and nobody liked it, but it was the only thing that kept the pricing honest. One thing that most people miss is that default risk is not static. It moves. A borrower with a solid track record for ten years can spike into danger very quickly if their industry gets hit by a regulation change or a major customer goes under. The lag in financial statement data makes this especially nasty. By the time you see the default rate climb in your reported numbers, the damage is often already done. This is why forward-looking indicators matter more than backward-looking ones, even though they are harder to build.
Another counter-intuitive point is that adding more data does not always improve accuracy. I worked on a project where we layered in hundreds of alternative data points. Social media activity, purchase history patterns, weather data for their region. The model accuracy went down. The extra variables introduced noise that overwhelmed the signal from the core financial ratios. Sometimes less data is the correct engineering decision. You have to keep the feature set tight and understand why each variable is in there. There are several approaches people use to estimate these probabilities. Structural models like the Merton model treat a company as having a certain distance to default based on its asset value and debt levels. This is elegant mathematically. It works reasonably well for large publicly traded companies with liquid equity markets. It falls apart for small private businesses where asset values are opaque and equity is not traded. Scorecard models based on logistic regression are the workhorse for retail lending. They are interpretable, regulatory friendly, and boring. That boringness is exactly why they last decades in production. Machine learning methods have become common in the last few years. Gradient boosting, random forests, neural networks. They can pick up complex non-linear relationships that traditional scorecards miss. The tradeoff is interpretability and stability. A black box model might give you a slightly better Gini coefficient but it can also drift in ways you cannot easily diagnose. Regulators are increasingly scrutinizing these models, especially for consumer credit decisions. If you cannot explain why a loan was denied, you have a compliance problem regardless of how accurate the prediction is.
Get the Full Details
Practical implementation usually starts with defining the default event clearly. Is it ninety days past due? days? A restructuring? The definition changes everything about your output. I once inherited a model where the default definition was inconsistent across different data sources. One system used thirty days past due, another used the legal covenant breach date. Merging them produced garbage. The fix was to standardize on a single definition across all pipelines before running any analysis. It took two weeks of argument with the data engineering team. It saved us months of rework later. Stress testing is another area where simple explanations fall short. You need to know what happens when multiple economic variables move against you simultaneously. A recession, rising unemployment, falling housing prices, and tightening credit conditions. The correlations between these factors change during stress periods. Things that looked independent in normal times start moving together. This is the difference between a model that looks good in a report and a model that survives a downturn. The biggest limitation of any default risk framework is that it can only be as good as the assumptions underlying it. Models assume continuity. Markets do not always respect that assumption. Black swan events exist and they will break your model. I learned this the hard way during the early days of the pandemic when foot traffic data for retail borrowers became completely uninformative almost overnight. The models that survived were the ones with simple fallbacks and clear triggers for manual review, not the ones with the most sophisticated algorithms.
If you are building something for a small operation, start with a basic logistic regression on the core variables you actually understand. Payment history, debt to income, collateral coverage, industry sector. Get the fundamentals right before adding complexity. Document every assumption. Plan for the day when your data source changes and your model breaks. Keep the number of parameters low enough that you can explain them to someone who did not build the thing.
