← ds learning track
notes · 14

Survival Analysis & Time-to-Default

Binary default models answer whether a borrower defaults. Survival analysis adds the missing dimension: when. That time structure is central for lifetime PD, IFRS 9 staging logic, and any model that needs a full default term structure instead of a single 12-month probability.

Start with censoring and the three core functions, then move into Kaplan-Meier, hazard shapes, Cox proportional hazards, and finally IFRS 9 PD term structures. The goal is to connect intuitive time-to-event logic with real credit-risk use cases.

Three core functions

Survival function S(t)

The probability that the borrower survives beyond time t without defaulting.

It starts at 1 and declines over time as default risk accumulates.

S(t) = P(T > t)

Hazard function h(t)

The instantaneous default intensity at time t, conditional on survival up to t.

This is the local speed of default risk.

h(t) = f(t) / S(t)

Cumulative hazard H(t)

The accumulation of hazard over time.

It links directly to survival and therefore to cumulative PD.

S(t) = e−H(t)
IFRS 9 link: lifetime PD is simply 1 − S(t). Marginal PD in year k is S(k−1) − S(k). That is the bridge from survival analysis to lifetime ECL mechanics. [oai_citation:2‡13-survival-analysis.html](sediment://file_00000000dfd8720abaefd47a314f29b7)

Why censoring changes everything

The core problem

In credit data, many borrowers do not default during the observation window. You know they survived at least until a certain date, but not what happens afterwards.

That is right-censoring. Ignoring it means losing valid time information and biasing the survival estimate.

Dropping censored cases usually overstates default risk.

Why logistic regression is not enough

A simple default / non-default model treats a 1-year non-default and a 10-year non-default identically.

Survival analysis uses the time dimension properly: how long the borrower stayed alive in the portfolio before default or censoring.

Credit reality: maturity, prepayment, observation cut-off, and cure all create incomplete event histories. Survival models are designed for exactly that setting. [oai_citation:3‡13-survival-analysis.html](sediment://file_00000000dfd8720abaefd47a314f29b7)

A useful order for learning survival analysis

01

Start with survival and hazard intuition

Before formulas, understand that one tells you how much survives and the other tells you how fast risk is arriving.

02

Then understand censoring

Most of the value of survival models appears precisely because some outcomes are incomplete at the observation date.

03

Then move to non-parametric structure

Kaplan-Meier gives the empirical survival curve before you start imposing functional assumptions.

04

Then add covariates and IFRS 9 logic

Once the time structure is clear, Cox PH and lifetime PD term structures become much easier to interpret.

Kaplan-Meier survival curve

Kaplan-Meier is the cleanest non-parametric estimator of S(t). Event times create downward steps. Censored observations stay in the risk set until they leave, but do not generate a drop themselves.

Hazard shape determines the PD term structure

Different portfolios do not share the same time profile of risk. Some have nearly constant hazard, some season downward, some rise with age, and some show a hump or bathtub structure.

Cox PH — adding borrower covariates

Cox PH allows covariates to shift the hazard up or down without forcing a specific baseline hazard shape. That makes it one of the most practical bridges between classical survival analysis and modern lifetime PD modelling.

From survival to lifetime ECL term structures

This is the practical endpoint for many credit-risk use cases. The survival curve turns into cumulative PD, marginal PD, and forward PD — exactly the quantities that drive lifetime ECL and stage-dependent measurement.

PD term structure — cumulative, marginal, forward

Cumulative PD Marginal PD Forward PD

Parameters

12-month PD (%)1.0
Hazard trend (k)1.0
Maturity (years)10

PD term table

Three PD concepts: cumulative PD = 1 − S(t), marginal PD = S(t−1) − S(t), forward PD = 1 − S(t)/S(t−1). These are related but not interchangeable. [oai_citation:8‡13-survival-analysis.html](sediment://file_00000000dfd8720abaefd47a314f29b7)
IFRS 9 staging: Stage 1 is mainly a 12-month view. Stage 2 and 3 require lifetime logic. That is why term structures matter far beyond academic survival modelling. [oai_citation:9‡13-survival-analysis.html](sediment://file_00000000dfd8720abaefd47a314f29b7)

Survival models compared

Model Type Hazard shape Handles censoring? Typical use
Kaplan-MeierNon-parametricEmpiricalYesExploratory survival estimation
Nelson-AalenNon-parametricEmpiricalYesCumulative hazard estimation
ExponentialParametricConstantYesSimple lifetime PD structures
WeibullParametricMonotone ↑ / ↓YesAgeing or seasoning effects
Cox PHSemi-parametricFlexible baselineYesLifetime PD with covariates
AFT / Log-logisticParametricMore flexibleYesWhen PH is doubtful
Discrete-time hazardSemi-parametricPeriod-basedYesMonthly / quarterly default modelling

Concepts every validator should keep

PH assumption

Proportional hazards may fail

If macro or behavioural effects change over time, a simple Cox PH assumption can be too rigid.

competing risks

Default is not the only exit

Prepayment, maturity, cure, and restructuring all compete with default and can distort naive lifetime estimates if ignored.

TTC vs PIT

Term structures depend on philosophy

TTC curves reflect long-run average behaviour. PIT curves absorb current and forward-looking macro conditions.

c-index

AUC has a survival analogue

The concordance index plays the role of discrimination summary for censored time-to-event models.

multi-horizon calibration

Calibration is not just 12-month

A survival-based lifetime PD model should be checked across multiple horizons, not only at the first year.

multi-state logic

Cure matters too

In some portfolios, default is not an absorbing endpoint. Cure and re-default can materially change the expected loss path.

What to leave this page with

Survival analysis adds the missing question to default modelling: not only whether default happens, but when.

The useful order is: first understand survival, hazard, and censoring; then read the empirical Kaplan-Meier curve; then interpret hazard shapes; then add covariates through Cox PH; then translate the result into IFRS 9 lifetime PD structures.

Once that structure is clear, lifetime PD stops looking like an abstract regulatory requirement and starts looking like a natural time-to-event model.