Probability of default, loss given default, and exposure at default are the three numbers every credit risk model is built on, and most people who use them cannot explain what they actually measure. That gap matters more than any model choice, because a number nobody can interpret is a number nobody can challenge. Working through a single loan to see how the three inputs combine into expected loss and, from there, into bank capital is the right way to close it. The mechanics are less intimidating than the acronyms suggest, and the same discipline of tracing a metric back to its source applies when you monitor agent behavior by tracking anomalies in execution history or ask why classic data science skills still shape smarter AI evaluations. In each case, the value comes from understanding what the measurement represents, not from the sophistication of the tool producing it.
Here is the practical takeaway: expected loss is a product, not a prophecy. Probability of default tells you how often borrowers fail, loss given default tells you how much is lost when they do, and exposure at default tells you how much was outstanding at that moment. Multiply them and you get an average loss across a portfolio. That average says nothing about whether any single loan will fail, and treating it as a prediction is where risk reporting quietly goes wrong. The three inputs also feed regulatory capital calculations, which means a small shift in an assumption about recovery rates can move a bank's capital requirement without any change in the underlying book. Anyone presenting these figures should be able to say which input moved and why.
The harder problem is judgment. Default probabilities come from historical data and statistical models, but the economy does not repeat itself on schedule. A model calibrated on a benign decade will understate stress, and the people closest to the portfolio often know that before the numbers do. Good risk practice keeps a human in the loop who can ask whether the assumptions still describe the world. That is the same principle behind letting AI handle dashboard visuals while you keep control of the business logic: automation handles the computation, people retain authority over the meaning.
Watch the assumptions, not the outputs. The next time a credit risk report crosses your desk, ask which of the three inputs changed since the last reporting period and what evidence drove it. If nobody can answer, the model is running the bank's understanding of risk rather than informing it.