Credit risk simulator: from the static scorecard to dynamic stress testing
In Latin American banking, fintech, and credit unions, prudential supervision no longer accepts a static bureau-driven scorecard. Mexico's CNBV, Peru's SBS, Colombia's SFC, Bolivia's SMV/ASFI, and Chile's CMF require internal models with quarterly backtesting, vintage segmentation, and stress testing under adverse scenarios aligned to Basel III. Annual expected loss is no longer reported as a historical average; it is decomposed, stressed, and reconciled against the regulatory capital estimate. Risk teams still running a master Excel with annual refreshes are years behind the minimum standard regulators and boards expect.
A modern credit risk simulator operates on the foundational equation of IRB Advanced:
Expected Loss (EL) = PD x LGD x EAD
where PD (probability of default) is the probability that a borrower will default within 12 months, LGD (loss given default) is the unrecoverable share after liquidating collateral and collecting, and EAD (exposure at default) is the expected balance at the moment of default. For revolving products (credit card, SMB line, payroll-deduct loan with a cap), EAD incorporates a CCF (credit conversion factor) that translates unused availability into projected exposure. In mortgage and auto lending, EAD includes outstanding principal plus accrued interest and pending fees at the cut-off date for default.
Numeric example: Mexican consumer portfolio
Assume a non-bank lender with a USD 29.4M consumer portfolio spread across 25,000 loans. With an average PD of 4.5%, LGD of 65%, and EAD equal to the outstanding balance, annual expected loss is:
EL = 0.045 x 0.65 x 29,400,000 = USD 860,000
That is 2.93% of the portfolio. If provisions sit at 3.0%, you are barely covering the mean - zero cushion for an adverse scenario. A stress test that raises PD to 8% (moderate recession, level observed in 2009 and partially in 2020) and LGD to 75% (lower recovery due to sectoral unemployment) lifts expected loss to USD 1.76M, 6% of the portfolio. The USD 880K gap against provisions is your capital shortfall - the metric CNBV and the risk committee track to trigger corrective actions before the problem moves from accounting to reputation.
If you further refine EAD with an 85% CCF on revolving lines (Basel II Foundation standard), the number rises another 3-5% on books with material revolving exposure. Moving from naive to calibrated typically adds 5-8 basis points to reported EL.
RAROC and risk-adjusted pricing
Risk-adjusted return on capital (RAROC) is the bridge between origination and treasury. It is calculated as:
RAROC = (Net income - EL - operating costs) / Economic capital
If RAROC for a segment falls below the cost of capital (typically 15-18% for LatAm fintech, 12-14% for traditional banking), that segment destroys value even if it sells. The simulator lets you evaluate each origination policy by RAROC before signing off on commercial pricing, avoiding the classic error of scaling placement while subsidizing growth with expected loss. Finance and credit align on the same number, one that is defensible before regulators and the audit committee.
Economic capital calculation requires defining a confidence interval (99% standard, 99.9% for systemic institutions), assumed default correlation, and an alpha multiplier on unexpected loss. The simulator lets you set those parameters on a single screen instead of rebuilding them from scratch in a spreadsheet every planning cycle.
Credit VaR and economic capital
Beyond expected loss, 99.9% credit VaR measures the maximum plausible loss over a 12-month horizon at that confidence level. The gap between VaR and EL is unexpected loss (UL), which Basel requires banks to cover with regulatory capital. For concentrated portfolios (above 20% in one sector or 5% in a single borrower), UL spikes due to default correlation - the sectoral contagion that shattered mortgage portfolios in 2008 and several LatAm fintechs in 2023 when tourism and delivery contracted together.
Credit VaR math is simulated with Monte Carlo over correlated defaults. With 10,000 trajectories, the 99.9 percentile gives you the loss in the worst 0.1% of cases - scenarios that mathematically occur every 1,000 years but appear every 20-30 years in real credit cycles due to fat tails. That is the figure the board needs to decide on additional capitalization, exposure limits, and reinsurance or portfolio securitization.
Backtesting and vintages
A PD model is not validated by looking at total default; it is validated by vintage. You segment the portfolio by origination month and track the cumulative default curve at 3, 6, 12, 18, and 24 months. If the June 2025 origination vintage shows 12-month cumulative default at 7%, but your model projected 4%, you have a 75% calibration drift - your scoring is broken or your target market has shifted. The simulator runs the same exercise on the inputs you define, with no data warehouse or SQL required.
Vintage analysis also exposes roll rate - the speed at which a loan migrates from 0-30 bucket to 30-60, to 60-90, and finally to write-off. A 30->60 roll rate of 60% means 6 out of 10 loans entering early delinquency end up in advanced delinquency. In healthy portfolios that number sits below 35%. The simulator loads your buckets and computes the implied Markov chain, projecting NPL formation and the next quarter's provisioning load.
Segmentation, scorecard, and policies
The advantage of a simulator over Excel appears when you want to evaluate 8 or 12 origination policies in parallel. Score cut-off changes, maximum amounts, maximum tenor by income level, flexible vs strict income verification, use of alternative data (transactional, telecom, psychometric) - each lever translates into deltas on placement, default, and RAROC. Instead of running 12 Excel tabs, you define policies as variants and the tool returns a comparison table. For the credit committee that is the difference between an intuition-based decision and one defended with quantitative evidence and stress scenarios.
In segments like SMB and microloans, where bureau data is limited or absent, behavioral scoring segmentation with at least 6 months of payment history becomes the backbone of the model. PD is recalibrated quarterly on a 24-month rolling window and subjected to out-of-time backtesting to avoid overfitting.
IFRS 9 provisions and forward-looking scenarios
Since 2018, IFRS 9 has mandated a shift from incurred-loss to forward-looking expected-loss provisions. The model requires estimating 12-month EL for Stage 1 portfolio (no significant deterioration) and lifetime EL for Stage 2 (significant deterioration) and Stage 3 (default). The simulator multiplies the complexity: you need stage-transition probabilities, projection of macro variables (GDP, unemployment, interest rate), and adjustments for statistical significance. Doing this in Excel for 40,000 loans is unviable; in the simulator it takes hours, not weeks.
Differentiation from spreadsheets
Excel solves an average. It does not solve inter-sector correlation, log-normal exposure distributions, or conditional dependence between PD and LGD (downturn LGD). A credit portfolio simulator with 10,000 Monte Carlo iterations on the EL = PD x LGD x EAD formula - each variable sampled from its distribution - delivers P50/P90/P99 percentiles that turn your board report from anecdote into quantitative evidence aligned with prudential supervision.
The simulator also produces regulator-ready artifacts: stress test tables with base, adverse, and severely adverse scenarios; EL decomposition by vintage and segment; sensitivities to rates and unemployment; and traceability of assumptions. Documentation that, in internal audit and supervisory visits, is the difference between a clean approval and a letter of observations.
Model governance and three lines of defense
A credit risk model does not live with the analytics team; it lives inside a model governance framework with three lines of defense. First line: the origination and collections team that uses the model daily. Second line: the independent Model Risk Management unit that challenges assumptions, replicates calculations, and approves or vetoes production use. Third line: internal audit that reviews the entire governance process. The simulator automatically generates the documentation - input datasets, assumptions, scenario results - that feeds the three lines without rebuilding it by hand each cycle.
In LatAm, Mexico's CNBV published in 2023 Article 242 bis of the Single Banking Circular, which mandates explicit model-governance documentation for all regulated institutions, aligned with SR 11-7 and the European TRIM framework. Non-compliance penalties are no longer hypothetical; Banxico and CNBV issued aggregate fines above USD 11.8M in 2024 for deficiencies in internal-model validation.
Conclusion
For risk teams in Latin American fintech, cooperatives, and mid-market banking, the simulator is the practical bridge between origination policy and the Basel III framework - without investing in SAS, Moody's CreditMetrics, Numerix, or a big-four consultant. Democratizing credit stress testing is part of the institutional maturity that separates a financial sector that learned from 2008 and 2020 from one that will discover the next crisis the hard way. The next time the board asks "what if rates rise 300 basis points and unemployment hits 7%?", having a quantitative answer in 10 minutes instead of 10 days changes the conversation and the reputation of the risk function inside the institution.