How Survival-Analysis Algorithms Model When Loans Default: Cox Hazards, Censoring, Time-Varying Covariates, Competing Risks and Calibration Diagnostics

Reader question: A one-year probability of default tells us whether default is expected within a fixed horizon. But what if we also need to model when default may occur, while correctly handling loans that repay, remain current or leave the dataset before default is observed?

This article owns the event-time modelling problem: loan histories + covariates + event and censoring rules → survival probabilities, hazard rates and time-dependent default probabilities that can be validated against observed timing.

It complements the separate probability-of-default calibration article, which focuses on mapping rating grades to one-year PD levels. Survival analysis adds the time dimension. It is also distinct from Merton structural credit risk, which infers a firm-value default signal from market equity and debt assumptions.

This is public mathematical and algorithmic education. It does not decide whether any person should receive credit, predict a named borrower’s default, or provide personalized financial advice.

The random variable is time to event

Let T denote the time from a defined starting point to default.

The cumulative distribution is:

F(t) = P(T ≤ t).

The survival function is:

S(t) = P(T > t) = 1 − F(t).

For credit applications, S(t) answers: what is the modelled probability that the borrower has not defaulted by time t under the stated event definition?

The cumulative default probability by time t is:

PD(0,t) = 1 − S(t).

The hazard rate asks a different question

The hazard rate describes the instantaneous event intensity among observations that have survived up to time t:

h(t) = lim_Δt→0 P(t ≤ T < t+Δt | T ≥ t) / Δt.

The cumulative hazard is:

H(t) = ∫₀^t h(u)du.

Survival and cumulative hazard are linked by:

S(t) = exp[−H(t)].

This identity is useful computationally because many survival models estimate a hazard structure and then convert it into survival probabilities.

Why survival analysis is useful for lending data

Loan datasets are rarely complete event datasets. At the end of an observation window, many loans are still performing. Some are repaid early. Some are transferred or otherwise disappear from observation.

A simple classifier that keeps only loans with known final outcomes can throw away useful information and bias the sample. Survival methods explicitly represent incomplete event observation through censoring.

Right censoring

If a loan has not defaulted by the last date it is observed, its event time is not known exactly. We only know:

T > C,

where C is the censoring time.

The record still contains information: the loan survived at least until C.

A survival dataset therefore often stores:

• start time;
• end or event time;
• event indicator;
• censoring reason;
• covariates measured at the appropriate observation time;
• event-definition version;
• loan or obligor identifier.

The censoring assumption is a weak link

Standard survival estimators rely on assumptions about censoring. Informally, after conditioning on the variables included in the model, censoring should not systematically hide event risk in a way the model ignores.

That assumption can fail in credit data. A borrower may refinance because risk improved, sell collateral because circumstances changed, or leave observation for reasons related to risk.

Therefore “not observed to default” is not automatically equivalent to “randomly censored.” Censoring reasons need their own diagnostics.

The Cox proportional-hazards model

A widely used semi-parametric survival model is the Cox proportional-hazards model:

h(t | x) = h₀(t) exp(βᵀx).

Here:

• h₀(t) is an unspecified baseline hazard;
• x is a vector of borrower, loan or macroeconomic covariates;
• β is a vector of coefficients.

The model is semi-parametric because it specifies how covariates multiply the hazard while leaving the baseline hazard shape unspecified.

Hazard ratios

If one coefficient is β_j, then increasing covariate x_j by one unit multiplies the hazard by:

HR = exp(β_j),

holding the other covariates fixed.

If β_j = 0.4:

HR = e^0.4 ≈ 1.49.

That means approximately 49% higher instantaneous hazard under the proportional-hazards interpretation. It does not mean the cumulative default probability is 49% higher at every horizon.

Partial likelihood: estimating relative risk without specifying the baseline first

The Cox model estimates β by comparing each observed event with the set of observations still at risk at that event time.

At an event time t_i, let R(t_i) be the risk set. The contribution for the borrower who defaults is proportional to:

exp(βᵀx_i) / Σ_j∈R(ti) exp(βᵀx_j).

Multiplying these contributions over event times yields the partial likelihood. The baseline hazard cancels from this coefficient-estimation step.

This separation is one reason the Cox model is attractive: it estimates relative hazard effects without forcing a parametric baseline shape such as exponential or Weibull.

Risk sets must be constructed correctly

The denominator is not “all loans in the database.” It contains loans still under observation and still at risk immediately before the event time.

If a loan has already defaulted, matured or legitimately left the risk set, it should not remain in later denominators.

Risk-set construction is therefore a high-value implementation test. An indexing error can corrupt every coefficient while producing output that still looks statistically plausible.

Tied event times

Loan systems often record events by day or month, so several defaults may share the same timestamp. The original Cox partial-likelihood derivation assumes distinct event times.

Common approximations for ties include the Breslow and Efron methods. The Efron approximation generally handles moderate ties more accurately than the simplest Breslow treatment while remaining computationally practical.

The tie method should be versioned because changing it can move fitted coefficients in datasets with many shared event dates.

From relative hazards to actual survival probabilities

After β is estimated, the model needs a baseline cumulative hazard H₀(t). A common estimator is the Breslow baseline cumulative hazard.

Then:

H(t | x) = H₀(t) exp(βᵀx),

and:

S(t | x) = exp[−H₀(t) exp(βᵀx)].

Equivalently:

S(t | x) = S₀(t)^exp(βᵀx).

A small conversion example

Suppose the baseline one-year survival probability is 95%, so:

S₀(1) = 0.95.

A borrower profile has a linear predictor βᵀx = 0.4, giving a hazard multiplier of about 1.49.

Then:

S(1|x) = 0.95^1.49 ≈ 0.926.

The corresponding cumulative one-year default probability is approximately:

1 − 0.926 = 7.4%.

This example shows why a hazard ratio is not itself a probability ratio. The conversion depends on the baseline survival curve and the horizon.

The proportional-hazards assumption

The Cox model assumes the hazard ratio between two covariate profiles is constant over time:

h(t|x_A) / h(t|x_B) = exp[βᵀ(x_A − x_B)].

If a variable has a strong effect immediately after origination but almost no effect several years later, a constant hazard ratio may be wrong.

Schoenfeld-residual diagnostics are a standard way to test whether coefficient effects show systematic time dependence. The R survival package’s cox.zph routine is one documented implementation of this test.

Time-varying covariates

Credit risk changes over a loan’s life. Potential covariates can include:

• delinquency state;
• current utilization;
• loan-to-value ratio;
• income or cash-flow updates;
• interest-rate resets;
• unemployment or macro indicators;
• remaining maturity.

A time-varying Cox model divides each observation into intervals where covariates are known and applies the appropriate state to each risk set.

The critical guardrail is temporal causality: a covariate value may enter the model only after it became observable. Using next month’s delinquency status to predict this month’s default creates future leakage.

Competing risks: prepayment can block the default event

A loan can terminate without default because it is prepaid, refinanced, sold, matured or otherwise settled.

If prepayment permanently removes the possibility of later default on that loan, it is a competing event, not merely missing data.

There are two common modelling questions:

• Cause-specific hazard: among loans currently event-free, what is the instantaneous hazard of default?
• Cumulative incidence: what is the probability of experiencing default by time t in a world where competing events can occur first?

Treating all prepayments as ordinary independent censoring can overstate or distort default incidence when prepayment is strongly related to borrower characteristics or market conditions.

Cause-specific hazard versus Fine–Gray style models

A cause-specific Cox model handles competing events by removing observations from the risk set after another event occurs and models the instantaneous cause-specific hazard.

A Fine–Gray subdistribution-hazard model targets cumulative incidence more directly by modifying how competing events remain represented in the risk set.

These models answer different questions. Their coefficients should not be treated as interchangeable.

Inputs and outputs

A production-shaped survival engine can require:

• loan or obligor identifier;
• observation start;
• event or censoring time;
• event type;
• default-definition version;
• static covariates;
• time-varying covariate history;
• macro data with publication timestamps;
• censoring reason;
• competing-event coding;
• tie-handling method;
• model specification and transformations;
• calibration horizon.

Outputs can include:

• coefficient estimates;
• hazard ratios;
• baseline cumulative hazard;
• survival curves;
• cumulative default probabilities by horizon;
• cause-specific cumulative incidence;
• standard errors and confidence intervals;
• concordance measures;
• calibration diagnostics;
• proportional-hazards tests;
• out-of-time validation results.

Evidence polarity

Evidence for confidence includes reproducible risk sets, stable event definitions, sensible coefficient signs, hazard ratios stable across reasonable model variants, survival probabilities calibrated to observed event timing, acceptable proportional-hazards diagnostics, consistent results on later time periods, and appropriate treatment of competing events.

Evidence against confidence includes systematic Schoenfeld-residual trends, implausible baseline hazard jumps caused by data-processing artifacts, large coefficient changes after small sample perturbations, censored observations concentrated in high-risk groups without explanation, materially different conclusions under competing-risk treatment, or strong deterioration on out-of-time cohorts.

Diagnostics that matter

• Risk-set audit: independently reconstruct who was at risk at sampled event times.
• Proportional-hazards test: examine Schoenfeld residuals and time interactions.
• Martingale/deviance residual review: look for functional-form misspecification and outlying observations.
• Concordance: test whether higher predicted risk tends to precede earlier observed events.
• Time-dependent discrimination: examine AUC or related measures by horizon rather than one pooled statistic.
• Calibration by horizon: compare predicted cumulative default probabilities with observed incidence at 6, 12, 24 or other relevant months.
• Brier score: measure probability error over time with suitable censoring treatment.
• Censoring sensitivity: compare results across censoring definitions and inverse-probability weighting where justified.
• Competing-risk test: compare naive survival estimates with cumulative-incidence analysis.
• Out-of-time validation: fit on earlier vintages and test on later vintages.
• Macro-regime stability: check whether coefficients or baseline hazards change after material economic shifts.

Counterexample: proportional hazards can fail

Suppose a high introductory interest rate strongly separates risk during the first six months but has little effect after refinancing or repricing.

A single constant hazard ratio averages two different regimes. The model can look acceptable in aggregate while being wrong at both ends of the timeline.

Time interactions, stratification, piecewise effects or alternative survival models can be more appropriate.

Counterexample: prepayment is not always harmless censoring

Imagine low-risk borrowers refinance quickly when rates fall while higher-risk borrowers remain in the portfolio. If refinanced loans are treated as ordinary independent censoring, the remaining risk set becomes selectively risky.

A naive default-survival curve can then misrepresent the probability of default in the original population.

Competing-risk modelling is one way to expose this mechanism.

Counterexample: a high C-index can coexist with bad probabilities

A model can order borrowers correctly from earlier to later default and still assign survival curves that are systematically too optimistic.

This mirrors the distinction between discrimination and calibration in one-year PD models. Ranking skill does not prove probability accuracy.

Counterexample: future information leakage can create spectacular but false performance

If a time-varying model uses a delinquency flag recorded after the default decision point, it can appear extremely predictive because the future is leaking into the past.

Every time-varying feature therefore needs an availability timestamp, not merely an accounting month.

Counterexample: sparse events make coefficients unstable

A segment with thousands of loans but only a handful of defaults does not necessarily support a large multivariable Cox model.

Separation, extreme coefficients and wide uncertainty can appear. Penalization, pooling, simpler models or external evidence may be needed.

Counterexample: macro regimes can move the baseline

A survival model fitted during one credit cycle can understate hazards during a later recession even if borrower ranking remains good.

Baseline-hazard stability, time interactions, macro covariates and vintage testing are therefore important update triggers.

Weak links in implementation

Ambiguous time zero. Origination, first draw, first payment and rating date are mixed.

Event-definition drift. Default criteria change without historical restatement.

Bad risk sets. Matured or already defaulted loans remain eligible at later event times.

Silent tie-method changes. Software upgrades change Efron/Breslow treatment.

Informative censoring. Exit depends on hidden credit condition.

Competing-event misclassification. Prepayment is treated as ordinary censoring without testing.

Temporal leakage. Updated borrower or macro information is used before its publication date.

Overparameterization. Too many covariates are fitted to too few events.

Probability conversion error. Hazard ratios are reported as PD multipliers.

Alternatives answer different questions

Discrete-time hazard models turn each loan-month or loan-quarter into a binary event opportunity and can use logistic or complementary-log-log links.

Parametric survival models such as exponential or Weibull impose a functional shape on the baseline hazard and can extrapolate more directly when the shape is credible.

Accelerated-failure-time models model covariate effects on survival time rather than proportional hazard.

Aalen additive-hazard models allow additive time-varying effects.

Random survival forests capture nonlinear interactions but require careful probability calibration and explainability checks.

Mixture-cure models can represent populations where a subgroup has effectively negligible event probability over the modelled horizon.

Simple one-year PD models may be preferable when the business question is purely fixed-horizon and timing adds little value.

How this connects to the surrounding knowledge estate

Survival analysis turns event timing into probabilities, while the PD calibration article focuses on long-run one-year grade calibration. Credit-rating transition matrices provide a discrete-state alternative for migration and default. IFRS 9 expected-credit-loss algorithms need horizon-dependent default information as one ingredient in lifetime loss estimation. Mortgage prepayment algorithms illustrate why prepayment can be economically important rather than ignorable censoring. Model validation supplies the independent challenge layer.

What would falsify confidence?

Confidence should be withdrawn when risk sets cannot be reproduced; proportional-hazards violations are material and unaddressed; censoring is clearly informative but treated as random; competing risks materially alter cumulative incidence; predicted survival curves repeatedly miss observed outcomes; coefficients change wildly under reasonable resampling; event definitions or time origins are inconsistent; or out-of-time cohorts show persistent calibration failure.

Verification and update triggers

Preserve the event definition, time origin, censoring rules, competing-event rules, data extract, feature timestamps, model formula, transformations, tie method, software version, baseline-hazard estimate and diagnostic outputs. Refit or revalidate after underwriting changes, servicing-system changes, default-definition changes, major refinancing waves, material macroeconomic regime shifts, portfolio acquisitions, feature-engineering changes, software upgrades, sustained proportional-hazards failure or out-of-time calibration drift.

Primary and high-quality references

• Marina Stepanova and Lyn Thomas, Survival Analysis Methods for Personal Loan Data, Operations Research, 2002.
• R Project, cox.zph documentation, for testing the proportional-hazards assumption using Schoenfeld residuals.
• Malvern Marimo and Charles Chimedza, Survival Analysis of Bank Loans and Credit Risk Prognosis, South African Statistical Journal, for application of survival and competing-risk methods to banking loans.
• Brian A. Ciochetti et al., Modeling the Termination of Commercial Mortgage Contracts: A Proportional Hazards Approach with Competing Risks, Real Estate Economics, for competing-risk event modelling.
• Basel Committee on Banking Supervision, CRE36 — IRB approach: minimum requirements to use IRB approach, for long-run one-year PD context and validation expectations.
• Federal Reserve Board and OCC, SR 11-7 — Guidance on Model Risk Management, for independent validation, outcomes analysis and model limitations.

Educational boundary: This article explains survival-analysis mathematics for event timing. It does not assess an individual borrower, make a lending decision, or provide personalized financial advice.