The model is not the decision
Every credit team in Latin America has lived some version of this conversation. The data science team presents a new model. The Gini coefficient is strong. The KS statistic looks good. The validation results are clean. Everyone agrees the model is better than the previous one. Then someone asks: what cutoff do we use? And the room gets quiet.
Because the model — however sophisticated — doesn't answer that question. The model produces a score. The score ranks customers by predicted probability of default. But the decision about where to draw the line between approve and reject is a different problem entirely. It's not a data science problem. It's an economic problem. And in most institutions across Chile, Peru, Argentina, Brazil, Mexico, and Colombia, it gets solved the same way it always has: with a spreadsheet, a meeting, and someone's best judgment. That judgment costs money. Often a lot of it.
What the cutoff actually determines
A credit score ranks customers by risk. Setting a cutoff at a given score level means approving everyone above that threshold and rejecting everyone below it. Move the cutoff up and you approve fewer customers, reduce defaults, but leave revenue on the table. Move it down and you approve more customers, capture more revenue, but increase losses.
This tradeoff is well understood. What's less understood is that the optimal cutoff is not a fixed number. It changes as a function of at least four variables that shift constantly: the cost of risk in the current portfolio, the price of the product, the institution's risk appetite, and the regulatory environment. A cutoff that was optimal six months ago may be significantly suboptimal today.
The gap between the score and the decision
Consider a typical credit portfolio in Colombia or Mexico. The model produces scores between 300 and 850. The institution sets a cutoff at 600. But within the approved population, there is enormous variation in profitability. A customer scoring 750 generates very different expected returns than a customer scoring 620. And within the rejected population, there are customers who would have been profitable approvals — customers whose behavioral profile makes them a better risk than their score suggests.
The economic cost of these two errors — approving unprofitable customers and rejecting profitable ones — compounds across hundreds of thousands of decisions. In a portfolio of 500,000 credit decisions per year, moving the effective cutoff by even a small amount in the right direction can mean tens of millions of dollars in improved profitability.
Why this problem is harder than it looks
The obvious response is: build a better model. This is correct but insufficient. A better model reduces both types of error — but it doesn't eliminate the cutoff problem. Even a perfect model still requires a cutoff decision. At what probability of default does the expected return become negative? That depends on the economics of the product, not on the model.
The cutoff problem is an optimization problem, not a prediction problem. What's needed is a layer that sits above the score and translates it into an economic decision — one that takes the predicted probability of default, combines it with current product economics, incorporates the institution's risk appetite and regulatory constraints, and computes the cutoff that maximizes expected value given all of those inputs simultaneously. A live optimization engine that answers: given everything we know right now about our costs, our margins, our risk appetite, and our regulatory environment, what is the cutoff that maximizes EBITDA?
What this looks like in practice
Response to macroeconomic shocks. When Argentina's peso devalued sharply in 2023, institutions with dynamic cutoff optimization adjusted their approval thresholds within days, before default rates in their portfolios began to rise. Institutions relying on static cutoffs and quarterly reviews adjusted months later — after the losses had already materialized.
Portfolio mix optimization. A bank in Chile running a dynamic economic decision layer can simultaneously optimize cutoffs across multiple products — personal loans, credit cards, BNPL, SME lending — taking into account the different economics of each product and the correlation of risk across segments.
Regulatory stress testing. Regulators across LATAM — the CMF in Chile, the SBS in Peru, the CNBV in Mexico — increasingly require institutions to demonstrate that their approval policies are economically justified. An institution that can show a regulator a live economic decision model, with explicit inputs and auditable outputs, is in a fundamentally stronger position than one presenting a static cutoff with no documented economic rationale.
The organizational challenge
In most financial institutions, the credit model is owned by the data science or risk modeling team. The cutoff decision is owned by the risk committee or the credit policy team. The product economics are owned by finance. These three groups typically don't sit in the same room, don't use the same tools, and don't operate on the same timelines.
Closing this gap requires a tool that speaks the language of all three groups simultaneously: the statistical language of the risk team, the financial language of the CFO, and the regulatory language of the compliance officer. A tool that takes the model output as input and produces an economic recommendation as output — with full auditability so every stakeholder can understand and defend the decision.
The score is the beginning, not the end
The credit score is one of the most powerful tools in financial services. But the score is not the decision. It never was. The decision is an economic optimization that takes the score as one input among several — and the quality of that optimization determines, more than almost any other factor, whether a credit portfolio is profitable or not.
Across Chile, Peru, Argentina, Brazil, Mexico, and Colombia, the institutions that will win the next decade of credit competition are not necessarily the ones with the best models. They're the ones with the best decisions — made faster, with more complete information, and with a clearer understanding of the economic implications of every cutoff choice. The model gets you to the starting line. The economic decision layer is how you win the race.