1Choose the kind of credit model a portfolio can support: application, behavioural, statistical rating or expert.
2Apply the regulatory default definition, 90 days past due on a material obligation or unlikeliness to pay, to a loan's history.
3Set an observation point and a performance window, and build development and out-of-time samples from a loan file.
4Fine and coarse class a characteristic and compute its weight of evidence and information value.
Why it matters
L3.3.04: Internal ratings and master scales gave RiskAtoms Bank plc a rating model with four weights and said real models are fitted to default data. Before anyone fits anything, someone must decide what a default is, when to start watching each loan and for how long. Watch the bank's 2023 personal loans for six months and 1.23% of them default; for two years and 4.95% do. Every model in this module inherits those choices, so this lesson makes them on the bank's own files.
Synthetic loan dataBooked figures standRiskAtoms Bank plc's lending files, to 31 December 2025
Real loan-level data is not public in a form anyone may reuse, so every loan and firm on this page is synthetic: generated by the course from fixed seeds, the same on every visit, and built to agree with RiskAtoms Bank plc's booked lending book of M3.3: Credit risk. No row is a real borrower.
The personal-loan file: 42,000 applicants in 12 quarterly vintages, 2023 to 2025, of whom 31,643 were accepted by the bank's older scorecard and four policy rules. A hidden “resilience” factor drives each applicant's characteristics and true risk; age has no direct effect on risk. How many loans default in each vintage and month is fixed by the booked vintage curves, so those match by construction and only who defaults is modelled. Declined applicants get a simulated outcome too, which no real bank sees.
The corporate and SME panel: 2,280 rated firms a year, 22,800 firm-years over 2016 to 2025, with 598 defaults equal, grade by grade and year by year, to the booked default history of L3.3.06: Probability of default. In each grade and year the firms with the weakest hidden quality default, so their ratios signal default more cleanly than real accounts would: expect high information values and AUCs. The economic cycle behind it is the course's own, not the UK's.
Which model?
The portfolio decides. An application scorecardApplication scorecard: a model used once, when someone applies, to decide whether to lend. It sees only what is known at the application. sees only what is known when someone applies: the bureau file, income, housing, the loan asked for. A behavioural scorecardBehavioural scorecard: a model re-run every month on existing customers, using how they have run the account (arrears, balances, limits). re-scores existing customers each month from how they run the account, and is usually much sharper (L3.3.14: Retail credit portfolios). For companies, a statistical rating model combines financial ratios with an analyst's view, as L3.3.04: Internal ratings and master scales's did. Where defaults are too rare to estimate from (banks, sovereigns, the strongest corporates), banks use expert modelsExpert model: a rating scheme whose weights are set by credit judgement rather than fitted, used where there are too few defaults to estimate from. whose weights are judged, and test them on whatever defaults exist. RiskAtoms Bank's files have all four situations: 1,383 defaults among 31,643 personal loans, 598 in 22,800 corporate firm-years, but only 2 in the 2,900 firm-years of grades 1 to 3, a low-default portfolioLow-default portfolio: a segment with so few defaults that a PD cannot be estimated from them reliably. that L6.3.05: Calibrating PDs returns to.
What is a default?
A model that predicts “default” must use the definition the bank will apply to it. For capital models that is the regulatory definition of defaultDefault (regulatory): the obligor is more than 90 days past due on a material credit obligation, or the bank judges it unlikely to pay in full without realising security (CRR Article 178).: more than 90 days past due on a material obligation, or unlikely to payUnlikeliness to pay (UTP): a default triggered by judgement on signs such as bankruptcy, a distressed restructuring or a specific provision, before or without 90 days past due.. In the personal-loan file, 1,296 of the 1,383 defaults came through 90 days past due and 87 (6.3%) through unlikeliness to pay in the first two months (fraud, sudden insolvency). After default, 242 loans (17.5%) cured: they paid their arrears and returned to performing. A cured loan still counts as a default for the PD; its low loss belongs to the LGD model (L6.3.08: LGD models). Sort these events:
Sort itDefault or not?
Each line is a fact about a borrower today. Which leg of the definition does it trigger, if any?
A card holder has missed four payments in a row; £640 is overdue on a £3,000 balance.
A company has filed for administration, though its loan payments are up to date.
A mortgage borrower is 45 days late after a payment holiday ended.
The bank agrees to cut a struggling firm's interest rate, reducing the loan's present value by 6%.
A personal loan is 120 days past due on £2 of interest left by a rounding error in the bank's own system; nothing else is owed late.
The bank sells a corporate loan at a 30% discount because of the borrower's credit problems.
When to look, and for how long
Each loan needs an observation pointObservation point: the date at which the model's inputs are read, for an application model the application date., when the inputs are read, and a performance windowPerformance window: the period after the observation point in which a default makes the loan a 'bad'. Twelve months matches a one-year PD. after it, in which a default makes it a “bad”. A longer window catches more of the defaults a loan will ever have, but leaves fewer recent loans with a complete outcome. For the 10,208 loans of the 2023Q1–2023Q4 vintages, all watched for at least two years:
Accepted loans of the 2023Q1 to 2023Q4 vintages, synthetic. A default after the window counts as a good.
Window (months)
Defaults
Default rate
Share of 24-month defaults
6
126
1.23%
25.0%
12
302
2.96%
59.8%
18
431
4.22%
85.3%
24
505
4.95%
100.0%
The application model here uses 12 months, which catches 60% of the two-year defaults and matches the one-year PD that capital needs. Lifetime PDs for IFRS 9 come from survival models instead (L6.3.07: Survival models for lifetime PD).
TimelineOne loan through the design
1 / 5
Mar 2023
Application
The observation point. Everything the model may use is read now: bureau score, income, debt payments, searches, the loan asked for. Nothing recorded later may enter, or the model leaks the future.
Samples
The development sampleDevelopment sample: the loans the model is fitted on. is the accepted loans of 2023Q1–2024Q1: 12,712 loans and 375 bads (2.95%). The out-of-time sampleOut-of-time sample: later loans kept aside to test the model on a period it has not seen. is 2024Q2–2025Q1: 11,183 loans and 466 bads (4.17%). Later vintages have less than a year of history. The out-of-time period is not just later; it is different. In 2024Q2 to 2024Q4 the bank loosened its rules (a lower cut-off and debt-to-income up to 60%), and acceptance rose from 73.1% to 82.2%; from 2025 an aggregator channel brought 2,120 loans, none of them in development. L6.3.04: Measuring discrimination and L6.3.05: Calibrating PDs test what that does to the model.
Only accepted loans have outcomes. The declined applicants, by reason:
All 42,000 applicants, 2023 to 2025, synthetic.
Declined because
Applicants
Income not declared
615
Bureau score below 550
161
Debt-to-income above limit
4,031
Score below cut-off
5,550
All declined
10,357
A model fitted on accepted loans alone learns about the people the old scorecard let in. How to correct for that, reject inference, is L6.3.03: Building a scorecard. The file has the usual flaws: no bureau score for 1.9% of applicants (thin files), undeclared time in job for 13.2% of the development sample, and 45 incomes typed in thousands or with an extra zero. Missing values keep a bin of their own; outliers land in the end bins.
Fine and coarse classing
Scorecards rarely use a characteristic raw. Fine classingFine classing: cutting a characteristic into many small bins (here up to 20 quantiles) to see how the bad rate moves. cuts it into many small bins and shows how the bad rate moves; coarse classingCoarse classing: merging fine bins into a few, each large enough to be stable, with a bad rate that moves in one direction. merges them into a few bins, each at least 5% of the sample, with a bad rate that rises or falls steadily. Each coarse bin then gets its weight of evidenceWeight of evidence (WoE): ln(share of all goods in a bin ÷ share of all bads in the bin). Positive means safer than average., and the characteristic its information valueInformation value (IV): the sum over bins of (share of goods − share of bads) × WoE; how well a characteristic separates goods from bads on its own.. Binning handles missing values and outliers, captures curves without a formula, and puts every characteristic on the same log-odds scale. The bureau score's coarse bins:
On the development sample, every characteristic of the personal-loan file:
12,712 development loans, bad = default within 12 months. Reading: Siddiqi's rule of thumb (below 0.02 useless, 0.1 to 0.3 medium, above 0.5 suspicious).
Characteristic
Fine bins
IV fine
IV coarse
Reading
Bureau score
21
0.643
0.589
suspicious
Gross income
20
0.443
0.423
strong
Debt-to-income
20
0.250
0.226
medium
Residential status
4
0.209
0.209
medium
Credit searches, last 6 months
7
0.179
0.153
medium
Employment
5
0.146
0.146
medium
Age band
6
0.094
0.094
weak
Months in current job
21
0.123
0.089
weak
Purpose
4
0.052
0.052
weak
Existing debt payments
15
0.086
0.034
weak
Loan amount
20
0.070
0.027
weak
Channel
3
0.025
0.025
weak
Term (months)
5
0.021
0.012
useless
The bureau score (0.589) is above the “suspicious” line. That line is a prompt to look for leakage, a characteristic recorded after the outcome. There is none here: the bureau score is read at application and is the closest proxy for the hidden resilience that drives default. Income (0.423) is strong, term (0.012) useless. Age band (0.094, highlighted) is weak but not zero, although the file gives age no direct effect on risk: it carries information through the characteristics it moves with, such as income and housing. The scorecard leaves it out, and L6.3.03: Building a scorecard checks that this is enough.
The corporate panel
The same tools on the 22,800 firm-years, with bad = default in the year and ratios from the previous year's accounts:
Corporate and SME panel 2016–2025, synthetic; numeric ratios in up to 20 fine bins merged to monotone with at least 2% each.
Characteristic
Bins
IV
Reading
Interest cover
17
2.560
suspicious
Net leverage
14
2.033
suspicious
EBITDA margin
13
1.000
suspicious
Qualitative score
5
0.534
suspicious
Current ratio
11
0.467
strong
Turnover
11
0.055
weak
Sector
8
0.019
useless
Read these with care
Interest cover's IV of 2.56 and net leverage's 2.03 are far above anything real company accounts give. In the synthetic panel the defaulters are exactly the firms with the weakest hidden quality in their grade and year, and the ratios are clean signals of it. Real ratios are noisier, late and sometimes managed. Sector (0.019) adds almost nothing. The ranking of the characteristics is instructive; the levels are not.
The formal version
Bads, bins and their evidence
For loan j with observation point tj and performance window W, and bins i = 1, …, k of one characteristic:
defaultj(s)=1{dpdj(s)>90 on a material amount}∨1{unlikely to pay at s}yj=1{maxtj<s≤tj+Wdefaultj(s)=1},W=12 monthsWoEi=lnbi/Bgi/G,IV=i=1∑k(Ggi−Bbi)WoEi≥0lnP(y=1∣bin i)P(y=0∣bin i)=lnBG+WoEi
gi, bi are the goods and bads in bin i and G, B their totals; a bin with no goods or no bads gets 0.5 added to both counts so its WoE is finite; missing values are always a bin of their own. Coarse bins: adjacent fine bins merged until the bad rate is monotone (pool-adjacent-violators), then any bin under the minimum share merged into its closer neighbour; the direction is the one with the larger IV. The IV is the symmetrised Kullback-Leibler divergence between the goods' and bads' distributions over the bins. Sources: Siddiqi (2017) for the classing practice; the IV rule of thumb is Siddiqi's, as quoted by Rojas, Alvarez and Rojas (2023); Thomas, Crook and Edelman (2017) for WoE and IV; CRR Article 178 for the default definition.
Try it
Pick a characteristic and watch the classing. At 20 fine bins and a 5% minimum the lab reproduces the IVs in the table: 0.589 for the bureau score. Force the bad rate to rise with the bureau score and the bins collapse to one, beside the missing values' own bin: the data refuse it. Raise the minimum share and the IV falls a little as bins merge; drop it to 1% and you keep more detail, with smaller and noisier bins. Try loan amount, which looks useful in 40 fine bins and gives most of it back once monotone.
Classing lab: from fine bins to a monotone WoE
Characteristic (development sample)
Fine bins (quantiles)
Bad rate must
Fine bins → coarse bins21 → 10
Bad rate across the binsfalling
IV, fine0.643
IV, coarse0.589
IV given up by merging0.055
Reading (rule of thumb)suspiciously strong (above 0.5)
A higher bureau score is safer, so the bad rate should fall across the bins. Missing values keep a bin of their own and are never merged. WoE = ln(share of goods ÷ share of bads) in the bin, so a positive WoE is safer than average; 12,712 development loans, bad = default within 12 months. Forcing the wrong direction collapses the bins: the data refuse it.
Worked example
The bureau score's weight of evidence
Two bins and the total
Step 1 of 4
1
Totals
The development sample has G = 12,337 goods and B = 375 bads. The bureau score has 10 coarse bins, the last for the 35 loans with no score.
Regulatory lens
The definition of default for IRB models is CRR Article 178: a default occurs when the bank considers the obligor unlikely to pay in full without realising security, or when the obligor is more than 90 days past due on a material credit obligation (Article 178(1)); Article 178(3) lists signs of unlikeliness to pay, among them non-accrued status, a specific credit adjustment, a sale at a material credit-related loss, a distressed restructuring and bankruptcy. In the EU the EBA's Guidelines on the application of the definition of default (EBA/GL/2016/07, applying from 1 January 2021) set the details: days past due count from the date a payment was due (paragraph 16), a distressed restructuring that lowers the obligation by more than 1% is a default (paragraphs 51 and 52), and a defaulted exposure returns to performing only after at least three months, a year after a distressed restructuring (paragraphs 71 and 72). The 2026 amending guidelines (EBA/GL/2026/05) kept the 1% threshold and the one-year period. In the UK, from 1 January 2027 Article 178 sits in the PRA Rulebook's Credit Risk: Internal Ratings Based Approach (CRR) Part, with the PRA's expectations in supervisory statement SS3/24 (January 2026). Data length is a rule too: PD estimates need an observation period of at least five years for at least one source, for corporates and retail alike (CRR Article 180(1)(h) and 180(2)(e)). The ten-year corporate panel meets it; the personal-loan file, with three years of vintages, falls short. The PD built on these choices is calibrated in L6.3.05: Calibrating PDs; the full list of IRB model requirements is L6.3.12: IRB model requirements.
Spark Atoms
✦Spark AtomScoring before computers+
In 1941 David Durand of the US National Bureau of Economic Research studied about 7,200 reports on instalment loans made by 37 lenders, from bank personal-loan departments to car and appliance finance firms, compared the good and bad loans, and built credit-rating formulae from them, with the theory of discriminant functions in an appendix. The design questions of this lesson, what counts as bad and which loans to compare, were already there.
Key takeaways
The portfolio sets the model: application and behavioural scorecards for retail, statistical rating models for companies with enough defaults, expert models for low-default segments.
Default is more than 90 days past due on a material obligation or unlikeliness to pay; in the personal-loan file 1,296 defaults came the first way and 87 the second.
A 12-month performance window catches 60% of two-year defaults and matches a one-year PD; development and out-of-time samples come from different vintages.
Only accepted loans have outcomes; declined applicants are the reject-inference problem.
Fine bins show the shape, coarse bins give stable monotone WoE; IV sums the evidence. A very high IV prompts a leakage check.
Synthetic ratios make corporate IVs implausibly high (interest cover 2.56); real data is noisier.
Reaction
Check what stuck
Question 1 · objective 4
A bin holds 400 of the sample's 10,000 goods and 40 of its 300 bads. What is its weight of evidence? (Two decimals.)
Question 2 · objective 2
A personal-loan customer is 20 days late with one payment when they file for bankruptcy. Under the regulatory definition, are they in default?
Question 3 · objective 3
For the 2023Q1–2023Q4 vintages, what share of the defaults seen within 24 months had already happened within 12 months? (Percent, one decimal.)
%
Question 4 · objective 4
The bureau score's IV on the development sample is 0.589, above the 0.5 that the rule of thumb calls "suspicious". What is the right reading here?
Question 5 · objective 4
Why does a scorecard builder merge fine bins until the bad rate rises or falls steadily?
Glossary
Application scorecard
Application scorecard: a model used once, when someone applies, to decide whether to lend. It sees only what is known at the application.
Behavioural scorecard
Behavioural scorecard: a model re-run every month on existing customers, using how they have run the account (arrears, balances, limits).
Expert model
Expert model: a rating scheme whose weights are set by credit judgement rather than fitted, used where there are too few defaults to estimate from.
Default (regulatory definition)
Default (regulatory): the obligor is more than 90 days past due on a material credit obligation, or the bank judges it unlikely to pay in full without realising security (CRR Article 178).
Unlikeliness to pay
Unlikeliness to pay (UTP): a default triggered by judgement on signs such as bankruptcy, a distressed restructuring or a specific provision, before or without 90 days past due.
Observation point
Observation point: the date at which the model's inputs are read, for an application model the application date.
Performance window
Performance window: the period after the observation point in which a default makes the loan a 'bad'. Twelve months matches a one-year PD.
Development sample
Development sample: the loans the model is fitted on.
Out-of-time sample
Out-of-time sample: later loans kept aside to test the model on a period it has not seen.
Fine classing
Fine classing: cutting a characteristic into many small bins (here up to 20 quantiles) to see how the bad rate moves.
Coarse classing
Coarse classing: merging fine bins into a few, each large enough to be stable, with a bad rate that moves in one direction.
Weight of evidence
Weight of evidence (WoE): ln(share of all goods in a bin ÷ share of all bads in the bin). Positive means safer than average.
Information value
Information value (IV): the sum over bins of (share of goods − share of bads) × WoE; how well a characteristic separates goods from bads on its own.
Low-default portfolio
Low-default portfolio: a segment with so few defaults that a PD cannot be estimated from them reliably.
More acronyms: DoD and UTP in the RiskAtoms glossary.
N. Siddiqi, Intelligent Credit Scoring: Building and Implementing Better Credit Risk Scorecards, 2nd edition, Wiley, 2017: the practitioner's guide to data, classing and scorecard development. Intermediate.
L. Thomas, J. Crook and D. Edelman, Credit Scoring and Its Applications, 2nd edition, SIAM, 2017. Advanced.
Origins
Regulation (EU) No 575/2013 (CRR), Articles 178 and 180, as onshored in the UK (legislation.gov.uk).
European Banking Authority, Guidelines on the application of the definition of default under Article 178 of Regulation (EU) No 575/2013, EBA/GL/2016/07, 2016; and amending Guidelines EBA/GL/2026/05, 2026.
Prudential Regulation Authority, SS3/24 Credit risk definition of default, January 2026.
N. Siddiqi, Intelligent Credit Scoring, 2nd edition, Wiley, 2017.
H. Rojas, C. Alvarez and N. Rojas, “Statistical hypothesis testing for information value (IV)”, arXiv:2309.13183, 2023.
D. Durand, Risk Elements in Consumer Instalment Financing, National Bureau of Economic Research, 1941.
L. Thomas, J. Crook and D. Edelman, Credit Scoring and Its Applications, 2nd edition, SIAM, 2017.
Watch this lesson
Video version
Prefer to watch? The same lesson as a 14-minute video will sit here, made from this page's script.