PD models.
Built from the defaults you can actually observe.

Rating and scorecard models for retail, SME and corporate portfolios, developed, calibrated and documented in Model Studio and designed around the EBA Guidelines on PD estimation.

Talk to a model specialist Decisions and tests →

Master scale · PD by rating gradeIllustrative

Regulatory basis

CRR · EBA GL on PD & LGD estimation · ECB Guide to internal models

Portfolios

Retail, SME, corporate, specialised lending

Outputs

Rating grades, master scale, long-run average PD, margin of conservatism

Used for

IRB capital, IFRS 9, stress testing

Approach

Risk differentiation first. Then quantification, with conservatism made explicit.

Default definition & sample

Align to the EBA definition of default and build a representative development sample.

Risk differentiation

Agents shortlist drivers and test combinations for discriminatory power and stability.

Calibration

Calibrate to the long-run average default rate, with the economic cycle reflected.

Margin of conservatism

Identify data and method deficiencies, quantify each, and document the resulting margin.

Data

What you need. And what happens if you don't have it.

Data	Needed	If missing or incomplete
Default history	Required · at least five years	Ask us about shorter histories
Obligor characteristics	Required	—
Financial statements	Helpful (non-retail)	Qualitative factors, with added conservatism
Payment behaviour	Helpful	Application data only
External ratings	Optional	Used as a benchmark

In Model Studio 0.1.0

Every choice is a decision for a person. Every test has its threshold on the page.

Decisions the agents propose

Each comes with a recommended option, the reason and any data gap. The model owner approves the set at the design gate.

Model approach
Which modelling approach fits the data?
Development sample
Which observation years form the development sample?
Hold-out sample
How is the model tested on data it was not estimated on?
Candidate drivers
Which fields may the agents consider as risk drivers?
Maximum number of drivers
How many drivers may the final model contain?
Segmentation
One model, or one per segment?
Master scale
Which master scale maps PDs to grades?
Margin of conservatism
At what confidence level is the general estimation error (category C) quantified?
PD input floor
Which PD floor applies to this portfolio?

Tests the engine runs

Shown next to each result. A test outside its threshold stays visible and becomes a limitation in the documentation.

Test	Group	Threshold
Gini · development	Discriminatory power	≥ 0.40 green · ≥ 0.30 amber
Gini · hold-out	Discriminatory power	≥ 0.40 green · ≥ 0.30 amber
Calibration tests per grade	Calibration	No grade with Jeffreys p < 0.05
Portfolio calibration	Calibration	Jeffreys p ≥ 0.05
Population stability (PSI)	Stability	≤ 0.1 green · ≤ 0.25 amber
Representativeness of the sample	Stability	Every driver PSI ≤ 0.25
Grade concentration (Herfindahl)	Stability	≤ 0.2 green
Override analysis	Use	Reported