Lifetime PD Is a Curve Through Time, Not 12-Month PD Multiplied by Maturity

One default event creates four distinct probability objects

Let τ represent default time over discrete periods t = 1, 2, …, T. P(τ = t) asks whether default occurs specifically in period t. P(τ ≤ t) asks whether default has occurred by t. Confusing them is the root of many lifetime-PD errors.

The sequence matters. Conditional risk is applied to the surviving population; survival shrinks; marginal PD allocates mutually exclusive default timing; cumulative PD adds those allocations. A reported 12-month PD may equal P(τ ≤ 1), and in a one-year discrete setting the first marginal PD. Practitioners must still document what each model output means rather than assume every “12-month PD” is constructed identically.

A constant 3% hazard does not produce 15% lifetime PD

The naïve 15% sum overstates the coherent result by 0.8734 percentage points because it applies 3% in later years to the original population rather than the survivors. At low hazards and short horizons, summation may look close; that is an approximation, not the probability architecture.

The marginal PDs sum to 14.1266%, exactly the unrounded CPD₅. They decline despite a flat 3% hazard because the population capable of first default becomes smaller each year.

Real risk has a shape, not merely a final value

Consider annual hazards of 2%, 3%, 4%, 3.5% and 3%. Risk builds through seasoning, peaks in year 3 and then moderates. The curve retains information that a 14.578% lifetime scalar discards.

The same lifetime PD can hide radically different timing

Curve A has hazards [5%, 4%, 2%, 1%, 1%]; Curve B reverses them to [1%, 1%, 2%, 4%, 5%]. Because both contain the same survival factors, each produces a lifetime PD of 12.4026%. Yet on an amortising €100k-equivalent exposure profile [100, 80, 60, 40, 20], 40% LGD and discount factors [0.96, 0.92, 0.88, 0.84, 0.80], Curve A produces an illustrative ECL of €3.601k versus €1.901k for Curve B.

Seasoning can create early peaks, gradual build, decline or a hump. Burn-out can also lower later hazards: if higher-risk borrowers default early, the surviving pool may be healthier even under unchanged macro conditions.

Lifetime ECL consumes marginal PD, not cumulative PD

MPD_t assigns each possible default to one period. Using CPD_t in every period—Σ CPD_t × LGD_t × EAD_t—repeatedly counts defaults already included in earlier cumulative probabilities. Lifetime PD is a useful end-of-horizon summary; the marginal curve is the integration object.

PD timing interacts with the other ECL dimensions. Amortising loans usually have EAD_t ↓, so early default carries more exposure. Revolving facilities may be drawn before default, making EAD flat or increasing and strengthening the future EAD / CCF research bridge. LGD and discount factors can also vary, so identical marginal-PD shapes need not produce identical losses.

Term structures can be developed through several methodological families

Direct horizon models must preserve CPD_t+1 ≥ CPD_t. Hazard models estimate h_t = f(X_t, t), naturally incorporating survival and time-varying covariates. Credit Vintage Analysis can supply CPD_v(m) shapes, but historical vintages may reflect different underwriting, channels and macro regimes.

A transition approach may use P^t to derive default-state probability under stability and first-order Markov assumptions: P(S_t+1 | S_t, S_t−1, …) = P(S_t+1 | S_t). Real credit risk often retains memory through previous delinquency, cure count, time in state, collections actions and seasoning. Roll Rate Analysis provides the connected diagnostic.

Where only a 12-month model exists, historical cumulative-default shapes, hazard scaling, segment curves or macro adjustments can extend it. Flat h_t = h₁ is a transparent benchmark, but not a default truth. Every extrapolation imports assumptions beyond the observed one-year model and must be validated as such.

Credit survival and exposure survival are different processes

The relevant horizon may depend on contractual maturity, behavioural maturity, prepayment, revolving behaviour and applicable extension options. A static contractual end date is not automatically the correct expected exposure period.

Prepayment or closure removes the exposure before default can occur on it. This creates competing risks: default, prepayment, maturity and closure are mutually relevant exits. Credit survival asks whether the borrower remains non-defaulted; exposure survival asks whether the facility remains capable of generating loss. A coherent ECL architecture must avoid projecting PD beyond an exposure that no longer exists.

Incomplete observation is uncertainty—not evidence of permanent survival

A recent loan observed for 18 months in a five-year product is right-censored: τ > T_obs is known, but its eventual five-year outcome is not. Treating every censored account as a permanent good biases lifetime risk downward. Survival methods incorporate partial time-at-risk more naturally than naïve final classification.

Long horizons require mature history. If T_obs < T_required, the unobserved tail must be estimated. A flat tail, decay assumption, external benchmark or long-run segment curve can each be defensible in context; none creates evidence where none exists. Governance should expose the assumption and sensitivity.

A long contractual tail need not dominate ECL when EAD has amortised, marginal risk is low or discounting is material. Conversely, sparse long-tail data in low-default portfolios can dominate model uncertainty. Preserve that uncertainty through alternatives and sensitivity rather than fabricated precision.

Future macro paths change the timing of default

Economically justified drivers may include unemployment, GDP, rates, inflation, property prices or sector conditions. A severe immediate shock followed by recovery can have the same five-year average unemployment as gradual deterioration but a different hazard path—and therefore different ECL.

Calculating one weighted PD curve first and then one deterministic loss can lose nonlinear interactions among PD, LGD and EAD. Scenario-specific loss followed by probability weighting may be more coherent where those interactions are material; this is a methodological question, not a universal implementation edict. Through-the-cycle curves provide smoother long-run risk, while IFRS 9 point-in-time architecture should respond to relevant current and expected conditions without collapsing every use case into one label.

SICR can depend on curve shape, not only a lifetime-PD ratio

Significant Increase in Credit Risk may compare LifetimePD^initial with LifetimePD^current. Yet two accounts can share the same ratio while one deteriorates mainly in years 1–2 and the other in years 4–5. The first may create more immediate management concern and a larger ECL increase on an amortising exposure.

Instead of comparing only CPD_T, analyse the full sets {CPD_t} or {h_t} across the remaining horizon. A curve-distance diagnostic can summarise where deterioration concentrates, but no one distance statistic should be treated as a universal SICR rule.

Validate level, shape and timing—not one final percentage

For mature vintages, compare PredictedCPD_t with ObservedCPD_t by horizon and PredictedMPD_t with observed period incidence. A model can match CPD_T while placing defaults in the wrong years. Vintage backtesting compares CPD^pred_v,t with CPD^obs_v,t; segment tests should cover product, risk grade, score, channel, customer and vintage only where default counts support inference.

Calibration Drift distinguishes a level shift from a shape shift. A simple intercept adjustment may repair level but cannot necessarily move defaults to the right periods. PD Model Monitoring and Credit Risk Model Validation extend the evidence to production and decision fitness.

Probability identities are powerful production controls

Enforce 0 ≤ CPD_t ≤ 1, CPD_t+1 ≥ CPD_t, MPD_t ≥ 0, S_t + CPD_t = 1 and Σ MPD_k = CPD_t. Calculate at controlled internal precision and round only for display. Version model, calibration date, macro scenario, curve, segment and effective date so historical ECL is reproducible.

A five-year loan connects every probability to ECL

Consider an original fictional amortising loan with a €100,000 opening exposure, the non-constant hazard curve above, evolving LGD and discount factors. Values are illustrative, calculated at full precision and rounded only for presentation.

A modest lifetime-PD change can conceal a material timing shock

Under an early-stress curve [4%, 5%, 3%, 2%, 1.5%], lifetime PD becomes 14.606%—only 0.027 percentage points above baseline. Yet ECL rises to €3,575.75, an increase of €647.48 or 22.1%, because marginal loss moves into years with higher EAD and less discounting. This is the practical difference between validating final level and validating shape.

Short-tenor and high-risk lending need monthly architecture

For six- or twelve-month consumer products, full lifetime can be short, data can mature quickly and hazard can be highly front-loaded. Monthly t = 1, …, 12 structures expose first-payment default, early seasoning, collections effects and rapid cure patterns that annual bank curves conceal.

When conditional monthly risk is high, Σh_t can diverge materially from 1 − ∏(1 − h_t), making survival mathematics especially important. High turnover and prepayment also mean exposure survival must be integrated explicitly. A copied annual long-duration curve is not proportionate sophistication; it is the wrong time unit.

For sparse long-tail or low-default portfolios, the opposite challenge applies: tail uncertainty may dominate. Governance should use materiality, sensitivity and clearly labelled extrapolation rather than pretending the tail is precisely observed.

A Lifetime PD agent should construct and challenge curves—not approve assumptions

A future Lifetime PD Term Structure Agent could assemble panel and vintage performance data; calculate empirical hazards and survival; derive marginal and cumulative PD; compare segments; identify immature tails; test alternative extrapolations; condition curves on approved macro scenarios; reconcile probability identities; compare predicted and realised curves; and detect level-versus-shape drift.

Its role is term-structure construction + validation + scenario analysis + monitoring support. It must not autonomously approve accounting policy, tail assumptions, macro scenarios or model calibration. Recurring value comes from recalibration, macro updates, monitoring and validation every reporting cycle.

The practitioner workflow is: Define Horizon → Estimate Conditional Risk → Apply Survival → Build Marginal Curve → Reconcile Cumulative PD → Condition on Scenario → Integrate with EAD/LGD → Validate Level & Shape → Monitor Drift.

Entimema's Credit Risk capability connects PD development, staging, validation and portfolio evidence. The CFO Function bridge connects default timing with reported allowance and impairment explanation. The parent IFRS 9 Expected Credit Loss architecture, SICR research, Default Definition, Credit Vintage Analysis and Roll Rate Analysis form the surrounding evidence system.