A Better Way To Model Provisions

Research evidence excerpt

A Better Way To Model Provisions

FoundationM

A Better Way to Model Provisions

Nubank's 1Q26 provision miss was not the asset quality break the market fears — it

exposed a bigger problem: the market is consistently getting provisions wrong.

Provision forecasting is inherently complex, yet most models still anchor on a simple cost

of risk assumption, often guided by management commentary, when provisions are

actually the output of multiple balance sheet dynamics. This matters because consensus

has repeatedly struggled to forecast provisions, particularly around first quarter

seasonality and credit inflection points. That is why we built a proprietary model that

captures the underlying drivers of provisions and gives us a more accurate, explainable

forecasting framework.

Our bottom-up, product-level roll-forward provision model decomposes NU's

provision line into the drivers that matter. Instead of starting with a cost of risk

assumption and multiplying it by loans, the model starts with the loan-loss allowance,

rolls forward exposures, stage migration, coverage, write-offs, recoveries, FX, macro, and

seasonality, and then derives the provision charge required to reach the appropriate

ending reserve. Cost of risk becomes the output, not the input.

The granularity required was substantial. For each product — credit cards and Loans to

Customers, including the secured/unsecured mix — we model exposure and allowances

across four risk buckets: Stage 1, Stage 2 relative trigger, Stage 2 absolute trigger, and

Stage 3. We then roll those balances forward in five steps: opening balances; volume

effects, including new originations, unutilized limits, and utilization changes; stage

migration through calibrated roll rates; coverage revaluation through severity drift and