Quick answer: banks forecast delinquency by treating a loan’s payment status as a state that can change over time. A current loan can become 30 days past due, cure back to current, roll forward to 60 or 90 days past due, pay off, restructure or eventually default/charge off depending on product and definition. Roll rates estimate how balances or accounts migrate from one delinquency bucket to another. Cure rates estimate how often delinquent accounts return to a performing state. Vintage curves compare cohorts originated at different times. More advanced hazard or transition models condition those movements on borrower characteristics and economic conditions. The result is a dynamic forecast of portfolio deterioration rather than one static delinquency percentage.

Thirty days past due is not an ending. It is a state with several possible exits.

Boundary: this article explains credit-risk mathematics and portfolio forecasting. It does not provide tactics for pressuring or harassing borrowers. Servicing, collections, accommodations and recoveries remain subject to contracts, law, consumer-protection requirements and institution-specific governance.

Why this belongs in mathematics

Delinquency forecasting uses Markov-style transitions, conditional probability, survival/hazard models, cohort analysis, time-series forecasting and state-dependent loss estimates. It also teaches why one observed percentage can hide motion. A portfolio with 5% delinquency can be improving rapidly if many accounts are curing, or deteriorating rapidly if current accounts are flowing into delinquency faster than old delinquencies resolve.

The New York Fed’s Household Debt and Credit reports explicitly track flows into delinquency, not only the stock already delinquent. Its August 2026 report showed why this distinction remains useful across mortgages, credit cards, auto loans and other household debt. See Household Debt Balances Decreased Slightly; Credit Card Delinquency Transition Rates Remained Steady.

1. Define the states before calculating transitions

A simple retail-loan state space might be:

Current;
1–29 days past due;
30–59 days past due;
60–89 days past due;
90+ days past due / serious delinquency;
Cured/current again;
Paid off;
Default/charge-off.

Different products and institutions use different status definitions. A stress-testing model may define mortgage default differently from accounting nonaccrual or operational collections status. The state definitions must match the model’s job.

The Federal Reserve’s mortgage stress-testing framework, for example, models transitions among current, impaired, default and paid-off states and allows impaired loans to cure back to current. See the Federal Reserve supervisory model descriptions.

2. Roll rate: how much moves to the next state?

Suppose S$10 million of credit-card balances are 30–59 days past due at the start of a month. At month-end:

S$3m cure to current or lower delinquency;
S$2m remain in 30–59;
S$4m roll to 60–89;
S$1m pay off, restructure or exit through another state.

The roll-forward rate to 60–89 is 4/10 = 40%. The simple cure/roll-back rate is 3/10 = 30%.

Repeated across every state, those flows form a transition matrix.

3. Transition matrices reveal the whole migration system

For a simplified state space [Current, Early Delinquency, Serious Delinquency, Default], a one-period matrix might look like:

From / To	Current	Early	Serious	Default
Current	0.94	0.05	0.00	0.01
Early	0.35	0.35	0.25	0.05
Serious	0.12	0.10	0.48	0.30
Default	0.00	0.00	0.00	1.00

Every row sums to 1. The matrix says not merely how many loans are delinquent but how quickly each state changes. Matrix powers can generate longer-horizon transitions if the underlying transition process is sufficiently stable—an assumption that must be tested rather than assumed.

This is structurally related to credit-rating migration models, but delinquency states often evolve faster and respond directly to payment behaviour.

4. Roll-rate forecasting is useful because it is simple

The OCC Credit Card Lending handbook notes that roll-rate analysis is widely used for short-horizon loss forecasting and collections workload planning. It is attractive because the calculation is transparent: observe how delinquency buckets migrated historically and apply those transition rates to today’s balances.

See the OCC Credit Card Lending handbook.

Transparency is a strength. It is also the source of the method’s main weakness: if underwriting, borrower mix or the economy changes, historical bucket-to-bucket motion can become stale before the headline delinquency rate shows it.

5. Why roll-rate power weakens at longer horizons

Roll rates work best when the near-term transition process is relatively stable. Over longer periods, too many other changes enter:

new originations change portfolio mix;
underwriting criteria tighten or loosen;
borrowers refinance or prepay;
macroeconomic conditions change;
account management changes;
collections or hardship programmes change;
legal and servicing rules change.

The OCC notes that delinquency-focused roll-rate models can lag deterioration in the large current bucket and may understate loss when portfolio quality is changing underneath the observed delinquency distribution.

6. Cure rate is not simply “good borrower” probability

A cure occurs when a delinquent loan returns to a defined performing state. Cure probability depends on:

how far delinquent the account already is;
borrower income and liquidity;
loan type;
payment shock that caused delinquency;
economic conditions;
availability and terms of lawful accommodations or restructuring;
collateral and refinancing options.

A 30-day-past-due mortgage and a 120-day-past-due unsecured loan should not inherit the same cure assumption merely because both are called “delinquent.” State depth matters.

7. Vintage curves reveal underwriting and cohort drift

A vintage groups loans by origination period—say Q1 2024, Q2 2024, and so on—and tracks performance as each cohort ages.

For each vintage, plot cumulative delinquency or loss against months-on-book. If the 2026 vintage reaches 5% serious delinquency at month 12 while the 2024 and 2025 vintages reached only 2%, the bank has evidence of cohort deterioration even if total portfolio delinquency is temporarily diluted by older strong loans.

Vintage analysis is especially useful after pricing, underwriting, marketing-channel or policy changes because it separates loan age from origination quality.

8. Hazard models condition transitions on the borrower and economy

Instead of assigning one roll rate to every loan in a bucket, a hazard or logistic transition model can estimate:

P(next state = j | current state, borrower variables, loan variables, macro variables).

Useful inputs can include payment history, utilisation, credit score, loan-to-value, interest-rate type, months-on-book, unemployment and property prices depending on product.

The Federal Reserve’s mortgage stress models use a system of conditional transition models rather than assuming one fixed roll rate for every mortgage. That allows economic conditions to change transition probabilities through the stress horizon.

9. Autoregressive delinquency models capture roll dynamics at portfolio level

Not every portfolio has granular loan-level data suitable for state-transition modelling. For some portfolios, the Federal Reserve has used delinquency and charge-off equations that include lagged delinquency/default rates, thereby capturing roll-rate dynamics implicitly at an aggregate level.

The difference is important:

loan-level transition model → follows individual states;
aggregate autoregressive model → predicts portfolio rates using their recent history and economic drivers.

Neither is universally superior. Data, use case, portfolio structure and validation evidence determine the appropriate resolution.

10. Cure and recovery are different

A loan that cures returns to performing payment status. A loan that defaults can still generate recoveries later through collateral, restructuring, sale, insurance or other lawful recovery processes.

A complete loss model therefore has at least two post-deterioration paths:

delinquency → cure → performing

delinquency → default → recovery timeline → final loss.

Recovery timing matters because receiving S$50,000 three years later is economically different from receiving it next month.

11. Workouts are decisions, not merely model states

For troubled commercial loans, banks can consider prudent workouts, restructurings, renewals or other alternatives consistent with applicable policy, accurate reporting and the aim of maximising recovery while controlling risk. The OCC describes problem-loan management as choosing among alternatives such as renewal, extension, restructuring or foreclosure depending on the circumstances.

See OCC Problem Loans.

A forecasting model should not assume that every delinquent loan follows a mechanical collections ladder unaffected by interventions. Accommodations and workouts can alter the state transition—but they need their own evidence and outcomes tracking.

12. Current portfolio flow matters more than the headline stock

In Q2 2026, the New York Fed reported aggregate household delinquency improving slightly to 4.7% of outstanding debt while flow into serious delinquency differed by product. Credit-card serious-delinquency flow remained elevated relative to some other categories, while mortgage flow edged higher year over year.

The educational lesson is not the specific US level. It is the measurement design: stock and flow answer different questions. A stable delinquency stock can hide high entry and high cure rates offsetting each other.

13. Creative-work lens: Groundhog Day and the importance of state transitions

Groundhog Day is not a credit source, but it offers a useful mathematical lens: observing the same calendar date repeatedly does not mean the system is unchanged if the path into and out of the state differs. A portfolio can report “5% delinquent” this month and next month while containing a completely different set of borrowers and a very different flow toward default or cure.

The creative work makes repeated-state thinking memorable. The credit model still needs account histories, lawful status definitions, macro evidence and realised outcomes.

14. The delinquency-forecast pipeline

Define mutually consistent payment states.
Reconcile status histories and balances.
Build monthly/quarterly transition counts and balances.
Calculate roll-forward, stay and cure rates.
Segment by product, vintage and borrower risk.
Build vintage curves and months-on-book views.
Add borrower and macroeconomic variables where justified.
Estimate transition/hazard or aggregate roll models.
Project default/charge-off timing.
Model recovery amount and timing separately.
Measure lawful accommodation/workout outcomes rather than assuming their effect.
Feed forecasts into staffing, loss allowance, pricing and stress testing as appropriate.
Backtest transitions, cures and recoveries.
Recalibrate when underwriting, economy or servicing practice changes.

15. Failure modes

Stock-only monitoring. Total delinquency is tracked without entry and cure flows.
Ordered-bucket assumption. Every account is assumed to pass through each delinquency stage mechanically.
Stale roll rates. Historical transitions survive major underwriting or macro changes unchallenged.
Vintage mixing. Strong old cohorts hide weak new originations.
Cure=recovery confusion. Returning to current payment status is mixed with post-default cash recovery.
Collections endogeneity. Changes in treatment strategy alter outcomes but the model attributes the difference entirely to borrower quality.
Terminal-state error. A model treats default or payoff as reversible without a defined reason, or makes a temporary delinquency state absorbing.
Consumer-context blindness. Operational optimisation ignores legal, hardship and fair-treatment constraints.

16. Diagnostics and falsifiers

What percentage of early delinquencies cure within three months?
Which bucket has the highest forward-roll probability?
Are recent vintages rolling faster than older vintages at the same age?
Does cure probability change sharply with unemployment or interest rates?
Are aggregate delinquency rates stable only because cure and entry are both high?
Do accounts skip delinquency stages before charge-off?
Do workout cohorts outperform comparable untreated or differently treated cohorts after accounting for selection?
Which assumption makes the nine-month forecast diverge most from realised loss?

Suppose someone claims, “The delinquency rate is unchanged, so credit quality is stable.” A falsifier is a transition matrix showing that current accounts are entering delinquency much faster while old delinquent accounts are curing or charging off at equally high rates. A stable stock can conceal unstable flows.

17. Verification and update triggers

backtest each transition probability by state and segment;
compare vintage curves at the same months-on-book;
reconcile charge-offs and recoveries to accounting records;
test cure assumptions after macro shocks;
separate model change from servicing-strategy change;
re-estimate after underwriting-policy changes;
monitor whether roll-rate errors concentrate in the current bucket;
keep simple transition matrices as challengers to more complex models.

Connections across the finance-and-banking algorithms lane

Revolving-credit utilisation — balance drawdown can accelerate before delinquency/default.
Expected credit loss — transition forecasts feed loss allowances under different accounting architectures.
Capital stress testing — stressed macro variables change delinquency and default flows.
Credit-rating migration — the same state-transition mathematics at a different credit scale.

Research anchors

The deeper lesson

Delinquency is a moving population. Roll rates reveal motion between payment states. Cure rates show that deterioration is not always one-way. Vintage curves show whether today’s new loans differ from yesterday’s. Hazard models add borrower and economic state. A strong bank therefore does not look only at how many loans are delinquent. It asks how they got there, where they are going next, which paths return to health, and which change in the world would make yesterday’s transition matrix stop working.

Educational note: This article explains public credit-risk and forecasting concepts. It is not debt-collection advice, consumer legal advice, servicing instructions or a recommendation about any individual borrower.

Small Group Tutorials

How Banks Forecast Loan Delinquency and Cure: Roll Rates, Transition Matrices, Vintage Curves, Cure Rates and Recovery Workflows