LAB · 03 / 09

Corporate credit risk

Fifteen years of delinquency by size and product, the policy-rate lag and a 6-month forecast with a stress scenario. Banco Central · time series · Ridge

Visual Light How it works ↓

How it works

Visual mode: shadows, ambient occlusion, depth of field and quality that adapts to your device. Light mode: no shadows or post-processing. Each bar is one cell of the data; the full analysis is right below.

Sheet of each analysis

Objective, why, target variable, predictors, method, metric, validation and whether the objective was reached.

Who defaults: households, companies, size

Objective: Compare household, corporate and company-size risk.
Why: A single credit policy for all companies errs both ways: too tight on large firms, too loose on small ones.
Target variable: Delinquency > 90 days (% of portfolio).
Predictors / explanatory variables: Segment (household, company, SME, large), month.
Method: Monthly SGS series; levels, peaks and size ratio.
Metric: % of portfolio; SME ÷ large ratio.
Validation: Banco Central definition used unchanged.
Objective reached?: Yes: SMEs show about 8 times large firms’ delinquency; households at a record.

Which products carry the risk

Objective: Find the corporate credit products that worsened most.
Why: Credit action is taken per product (limit, rate, collateral); a per-product alert is actionable, a general one is not.
Target variable: Current and 5-year maximum delinquency by product.
Predictors / explanatory variables: Free-funding product, month.
Method: Current level against its own worst month in 5 years.
Metric: % of portfolio; current = 5-year max.
Validation: Trade-finance products act as a control (low risk expected).
Objective reached?: Yes: working capital, asset purchase and secured overdraft at their 5-year worst.

Rates rise today, arrears come later

Objective: Measure how long the policy rate takes to reach corporate delinquency.
Why: Knowing the lag gives the warning time: limits can be adjusted before arrears appear, not after.
Target variable: 12-month change in corporate delinquency.
Predictors / explanatory variables: 12-month change in the policy rate, lagged 0 to 24 months.
Method: Cross-correlation on changes (avoids spurious trend correlation).
Metric: Pearson correlation by lag.
Validation: Curve with a single peak and decline on both sides, not an isolated point.
Objective reached?: Yes: 11 months, correlation 0.77.

6-month forecast and stress scenario

Objective: Forecast corporate delinquency 6 months ahead better than repeating the last value and measure a rate shock.
Why: Expected-loss provisioning (Res. 4,966 / IFRS 9) requires forward-looking scenarios; a simple, auditable model serves that better than an accurate, opaque one.
Target variable: Change in corporate delinquency over the next 6 months.
Predictors / explanatory variables: Current level, 6-month trend, policy rate lagged 11 months, rate change, IBC-Br, SME − large gap.
Method: Ridge regression (α = 1).
Metric: Mean absolute error (p.p.) and direction hit rate.
Validation: Rolling origin since 2016: each month trains only on targets already observed; 122 forecasts.
Objective reached?: Yes: 11% less error than repeating the last value, 79% direction hit rate.

Banco Central series

Delinquency over 90 days by client type, size and product, policy rate and IBC-Br, monthly since 2011 (SGS).

Validated forecast

Ridge regression 6 months ahead, validated with a rolling origin: each month the model only uses what was already known. Stress scenario with a policy rate 3 points higher.

Δinad(t+6) = f(nível, Selic(t−11), IBC-Br, …)

Corporate credit risk in data

Fifteen years of delinquency in the financial system, by client type, company size and product, crossed with the policy rate and economic activity. The underlying question is a corporate and middle-market bank’s: when will corporate credit deteriorate, in which products, and how much does a rate shock weigh on it.

Real-world use

Expected-loss provisioning (CMN Resolution 4,966, aligned with IFRS 9) and risk appetite: provisions must be forward-looking with macro scenarios, and the credit team needs to know in which product and client size to tighten or loosen.

3.35%corporate delinquency in 2026-08

5.98 × 0.72%SMEs × large companies

11 monthslag from policy rate to corporate delinquency

186months of data (2011–2026)

Who defaults: households, companies, size

QuestionDoes corporate credit risk behave like household risk? Does size matter?

Decision and whyI used the Banco Central definition: delinquency is the share of the portfolio with any instalment more than 90 days overdue. I split companies from households and, within companies, micro/small/medium from large ones, because risk profile and products differ.

ResultCorporate delinquency peaked at 4.06% in 2017-05 and is now 3.35%. Household delinquency is at the highest point of the series (6.03%). Among companies, size separates everything: SMEs at 5.98% versus 0.72% for large companies, about 8 times higher.

Source: Banco Central · SGS 21083, 21084, 27703, 27704

Analysis technical sheet

Objective: Compare household, corporate and company-size risk.
Why: A single credit policy for all companies errs both ways: too tight on large firms, too loose on small ones.
Target variable: Delinquency > 90 days (% of portfolio).
Predictors / explanatory variables: Segment (household, company, SME, large), month.
Method: Monthly SGS series; levels, peaks and size ratio.
Metric: % of portfolio; SME ÷ large ratio.
Validation: Banco Central definition used unchanged.
Objective reached?: Yes: SMEs show about 8 times large firms’ delinquency; households at a record.

Delinquency > 90 days (% of portfolio)

Households
SMEs
Companies (total)
Large companies

Which products carry the risk

QuestionWhich corporate credit products are riskiest today, and how much have they worsened?

Decision and whyI compared the current level with the worst month of the last 5 years for each free-funding product. A product sitting at its own worst level needs attention even if its absolute rate is low.

ResultOverdraft and credit cards concentrate the highest arrears, as expected for revolving credit. The more important signal is another one: working capital (5.98%), asset purchase and secured overdraft are at their worst level in 5 years. They are the everyday products of mid-sized companies. Trade finance (ACC, export finance) stays below 1%, partly because it is backed by the exports themselves.

Source: Banco Central · SGS 21087–21109

Analysis technical sheet

Objective: Find the corporate credit products that worsened most.
Why: Credit action is taken per product (limit, rate, collateral); a per-product alert is actionable, a general one is not.
Target variable: Current and 5-year maximum delinquency by product.
Predictors / explanatory variables: Free-funding product, month.
Method: Current level against its own worst month in 5 years.
Metric: % of portfolio; current = 5-year max.
Validation: Trade-finance products act as a control (low risk expected).
Objective reached?: Yes: working capital, asset purchase and secured overdraft at their 5-year worst.

Delinquency by product (companies, free funding)

current (2026-08)
worst month in 5 years

When both bars are the same length, the product is at its worst level in 5 years.

Rates rise today, arrears come later

QuestionHow long does a policy-rate hike take to show up in corporate delinquency?

Decision and whyI correlated the 12-month change in the policy rate, shifted by 0 to 24 months, with the 12-month change in corporate delinquency. Using changes instead of levels avoids the spurious correlation between two merely trending series.

ResultThe relationship is weak in the same month (0.24) and strongest with a 11-month lag (0.77). In practice: what the policy rate did over the last year is the best warning of what the corporate book will show over the next.

Source: Banco Central · SGS 4189 (Selic) e 21083

Analysis technical sheet

Objective: Measure how long the policy rate takes to reach corporate delinquency.
Why: Knowing the lag gives the warning time: limits can be adjusted before arrears appear, not after.
Target variable: 12-month change in corporate delinquency.
Predictors / explanatory variables: 12-month change in the policy rate, lagged 0 to 24 months.
Method: Cross-correlation on changes (avoids spurious trend correlation).
Metric: Pearson correlation by lag.
Validation: Curve with a single peak and decline on both sides, not an isolated point.
Objective reached?: Yes: 11 months, correlation 0.77.

Policy rate → corporate delinquency correlation by lag

Horizontal axis: months of lag.

6-month forecast and stress scenario

QuestionCan corporate delinquency 6 months ahead be forecast better than repeating today’s value?

Decision and whyRidge regression on the change over the next 6 months, using the current level, the recent trend, the policy rate lagged 11 months, the rate change, activity (IBC-Br) and the SME-to-large gap. Rolling-origin validation: at each month since 2016 the model only trains on what was known at that month. Baseline: repeat the last value, which is hard to beat for slow series.

ResultAcross 122 out-of-sample forecasts, mean error fell from 0.301 to 0.267 percentage points (11% lower), and the model called the direction of the move in 79% of cases. In the stress scenario, a sustained 3-point higher policy rate adds about 0.41 point to corporate delinquency within 6 months: on a R$ 10 bn book, roughly R$ 41 m more in arrears.

Source: Banco Central · SGS · analises/banco.py

Analysis technical sheet

Objective: Forecast corporate delinquency 6 months ahead better than repeating the last value and measure a rate shock.
Why: Expected-loss provisioning (Res. 4,966 / IFRS 9) requires forward-looking scenarios; a simple, auditable model serves that better than an accurate, opaque one.
Target variable: Change in corporate delinquency over the next 6 months.
Predictors / explanatory variables: Current level, 6-month trend, policy rate lagged 11 months, rate change, IBC-Br, SME − large gap.
Method: Ridge regression (α = 1).
Metric: Mean absolute error (p.p.) and direction hit rate.
Validation: Rolling origin since 2016: each month trains only on targets already observed; 122 forecasts.
Objective reached?: Yes: 11% less error than repeating the last value, 79% direction hit rate.

Corporate delinquency: actual × forecast made 6 months earlier

actual
model
repeat last value

11 %less error than repeating the last value

79 %direction hit rate (up/down)

What each dataset contributed

Delinquency by size: showed that "company" is not a single risk; SMEs and large firms need different policies.
Delinquency by product: flagged working capital and asset purchase at a 5-year worst, the most actionable warning.
Policy rate: gave the warning horizon, about a year, to adjust limits before arrears show up.
Together they fed a simple model validated without looking ahead, usable as a provisioning scenario.

Limits of this analysis

System-wide aggregates: they do not replace a bank’s own book, which has a different client and product mix.
Delinquency > 90 days is a lagging indicator; 15-to-90-day arrears would warn earlier but are not in these series.
The rate shock is a reading of the linear model, not a full stress test with second-order effects.

Analysis code: analises/banco.py