Probability of Default: Definition, Formula & Examples

Updated 12 Mar 2026Content was accurate at the time of publication

What is Probability of Default?

Probability of Default (PD) — also called default probability or credit default probability — is the likelihood that an obligor (borrower or counterparty) will fail to meet contractual debt obligations within a specified horizon, typically one year. PD is a forward-looking, per-obligor risk metric used by credit risk teams, investors and regulators. It differs from an observed default rate (a historical cohort frequency) because PD is an estimated probability for each exposure based on information available today.

Intuitively, PD answers: "Given what I know now about this borrower and the macro environment, what is the chance they will default within my chosen horizon?" You can estimate PD at the instrument level (loan, bond) or obligor level (company, household). PD is combined with Exposure at Default (EAD) and Loss Given Default to compute expected loss.

Why PD Matters: use cases

PD underpins most credit processes and risk measures:

Credit decisioning and pricing — PD helps set pricing spreads and lending limits reflecting expected credit losses; relevant for commercial products such as business loans and asset finance.
Expected Loss (EL) and provisioning — EL = PD × EAD × LGD; PD drives provisions and P&L recognition.
Capital calculation — under IRB approaches PD feeds regulatory capital models and stress testing.
Portfolio management — aggregate PD distributions inform concentration risk, scenario analysis and optimisation.
Counterparty limits and collateral management — PD guides exposure limits and margining.
Disclosure and governance — PD models are subject to validation and governance expectations.

PD also influences product-level decisions for lenders and secured portfolio risk management.

Core formula: PD in expected loss and capital calculations

The canonical expected loss formula is:

\text{Expected Loss (EL)} = \text{PD} \times \text{EAD} \times \text{LGD}

PD = probability of default (usually over 1 year).
EAD = exposure at default (amount outstanding at default, including off-balance items).
LGD = loss given default (1 − recovery rate), expressed as a proportion of EAD.

Example: PD = 2% (0.02), EAD = AUD 100,000 and LGD = 45% (0.45):

\text{EL} = 0.02 \times 100{,}000 \times 0.45 = \text{AUD }900

PD also enters regulatory capital formulas under IRB, where capital is calibrated to unexpected loss (UL) which depends on PD, LGD and asset correlation assumptions.

Types of PD: Point-in-Time (PIT) vs Through-the-Cycle (TTC)

PIT vs TTC is central to model design and use:

Point-in-Time (PIT) Estimates conditioned on current macro and borrower signals; cyclical and responsive to near-term conditions. Best for pricing, provisioning and forward-looking stress tests, but volatile and requires frequent recalibration and macro overlays.

Through-the-Cycle (TTC) Smoothed estimates over the credit cycle, representing an average default likelihood. Stable and useful for long-term capital allocation and comparability, but can understate near-term risk and is less useful for provisioning and pricing.

Conversion & calibration: You can derive PIT PDs from TTC models using a cycle factor or macro model (e.g., forecasted GDP/unemployment). Model governance must document whether PDs are reported PIT or TTC and how conversions are performed.

Common estimation approaches

Three broad PD estimation approaches exist:

Statistical (data-driven) models: logistic regression, survival/hazard models, and machine learning classifiers using borrower and loan features.
Structural models: connect firm value dynamics to a default barrier (suitable for listed corporates).
Market-implied measures: derive implied default probabilities from traded instruments such as CDS spreads and bond yields.

Each approach has trade-offs in data needs, interpretability and use-case fit. Statistical models are flexible and widely used; structural models suit firms with observable market data; market-implied PDs are timely but can be noisy.

Statistical models: logistic regression, survival/hazard and machine learning

Statistical methods are pervasive for retail and SME lending and for obligors without liquid equity markets.

Logistic regression Binary outcome: default within horizon (1) or survival (0). The model specification is logit(PD) = β₀ + β₁X₁ + … + βₖXₖ. Score-to-PD conversion: PD = 1 / (1 + exp(−score)). Typical predictors include financial ratios (interest cover, leverage), payment history, utilisation, borrower age, industry and macro variables. Calibration requires choosing a training window and handling class imbalance (defaults are rare) with sampling or penalisation.

Survival / hazard models Useful for modelling time-to-default and handling censoring (prepayments, cures). They estimate an instantaneous hazard λ(t|X) and convert it to a cumulative PD over the required horizon. Use when time at risk varies across accounts or when censoring is material.

Machine learning Tree-based methods and boosting often improve ranking but can miscalibrate probabilities. Convert ML scores to PDs via Platt scaling (sigmoid), isotonic regression or binning calibration. Use explanation tools (SHAP, partial dependence) to retain interpretability for governance.

Practical tips:

Separate discrimination (ranking — e.g., AUC) from calibration (do predicted PDs match observed defaults).
Document variable selection, assumptions and data lineage.

Structural models (Merton-type)

Structural models relate a firm's default risk to economic fundamentals and market prices. They leverage equity prices and balance-sheet information to estimate a firm's distance-to-default. Strengths include connecting fundamentals to default risk and using market signals. Limitations include the requirement for liquid market data, reliance on simplifying assumptions, and potential to miss defaults occurring between maturity dates.

Structural models are complementary to statistical and market-implied approaches, especially for large corporates with liquid equity.

Market-implied PDs (from CDS spreads and bond prices)

Market-implied PDs use traded instruments that reflect credit risk. A simple approximation under a constant hazard rate h and assumed recovery R is:

s \approx (1 - R)\,h \quad\Rightarrow\quad h \approx \frac{s}{1-R}

and the 1-year PD is

\text{PD}_{1y} = 1 - e^{-h} = 1 - \exp\left(-\frac{s}{1-R}\right).

Caveats: CDS spreads include liquidity, counterparty and funding premia. Recovery R is uncertain; different R values materially change PD. Bond-implied PDs require stripping risk-free and liquidity premia.

For market context and methodologies see the Reserve Bank of Australia (RBA) at https://www.rba.gov.au/ and the Bank for International Settlements (BIS) at https://www.bis.org/.

Mapping ratings to PDs and transition matrices

Ratings-to-PD mappings use historical default studies (Moody's, S&P). Transition matrices show empirical probabilities of moving between rating buckets over time (e.g., 1-year transitions).

Uses include:

Portfolio forecasting (roll-forwards by rating)
Stress testing (apply stressed PDs)
Credit migration analysis for pricing and accounting (IFRS 9)

When using mappings, perform vintage analysis and update tables to reflect current default environments.

Worked examples

Example 1 — Loan-level PD (logistic)

Model: logit(PD) = −3.0 + 0.8·(leverage) − 0.5·(interest cover) + 0.6·(payment delinquencies)

Borrower data: leverage = 2.0, interest cover = 4.0, delinquencies = 1

Calculation:

Score = −3.0 + 0.8(2.0) − 0.5(4.0) + 0.6(1) = −2.8
PD = 1 / (1 + e^(2.8)) ≈ 0.0575 → 5.75%

Example 2 — CDS-implied PD

Input data: 1-year CDS spread s = 300 bps = 0.03; assume recovery R = 40% (0.40)

Calculation:

Hazard h = s / (1 − R) = 0.03 / 0.60 = 0.05
1-year PD = 1 − exp(−0.05) ≈ 4.88%

Spreadsheet steps: For logistic PD, implement score formula then PD = 1 / (1 + EXP(−score)). For CDS PD, input spread and recovery; hazard = spread / (1 − recovery); PD = 1 − EXP(−hazard × horizon).

Data requirements, quality and common pitfalls

Key data needs include sufficient history of defaults and survivals (including censored observations), granular borrower characteristics (financials, payment records, behavioural data), macro indicators for PIT calibration (GDP growth, unemployment, house prices), and market data for structural/market-implied models (equity prices, CDS spreads).

Common pitfalls to avoid:

Small sample bias: rare defaults require pooling or hierarchical models.
Survivor bias: excluding failed entities understates PDs.
Censoring: ignore prepayment and cure patterns at your peril — use survival methods that handle censoring.
Leakage: including forward-looking indicators unavailable at decision time biases validation.
Calibration drift: economic regimes change — recalibrate and document triggers.

Model validation, backtesting and governance

Validation checklist:

Discrimination: Test AUC (ROC area) and KS statistic.

Calibration: Create calibration plots (predicted PD bins vs observed defaults) and calculate Brier score:

\text{Brier} = \frac{1}{N}\sum_{i=1}^N(\hat{p}_i - y_i)^2

where yi is the 0/1 default outcome.

Backtesting PDs: Perform unconditional backtest (compare average predicted PD to observed default rate) and conditional backtest (test calibration by score band or rating). Use appropriate statistical tests but account for low default frequency.

Stress & sensitivity testing: Apply macro stress scenarios and re-run PDs. Test sensitivity to recovery assumptions for market-implied PDs and to model choices.

Governance & frequency: Document purpose, inputs, assumptions and limitations. Recalibrate PIT PDs at least quarterly or on material performance drift; TTC PDs less frequently. Maintain independent validation, version control and audit trail in line with supervisory expectations.

Step	Purpose
Data lineage check	Verify completeness and censorship handling
Discrimination tests (AUC, KS)	Assess rank ordering
Calibration plots & Brier score	Assess probability accuracy
Backtesting by vintage	Track out-of-time performance
Sensitivity to recovery/macro	Quantify fragility
Governance documentation	Ensure auditability and controls

Regulatory & industry context

Under the Basel framework, IRB approaches require institutions to estimate PD, LGD and EAD with robust models and governance. PD estimates feed risk-weighted assets and capital via functions that consider default correlation and PD level. For Basel material see the BIS at https://www.bis.org/.

Regulatory expectations in Australia align with Basel principles. Maintain credible model governance, independent validation and documented calibration. Model outputs used for regulatory capital are subject to supervisor review; APRA publishes prudential guidance and thematic reviews at https://www.apra.gov.au/. ASIC covers conduct and disclosure where credit decisioning affects consumers at https://asic.gov.au/. The Reserve Bank provides credit and financial stability data that inform macro overlays at https://www.rba.gov.au/. Relevant legislation and insolvency frameworks are published on the Australian Government legislation site at https://www.legislation.gov.au/. For tax and administrative context that can affect credit outcomes see the Australian Taxation Office at https://www.ato.gov.au/.

When preparing regulatory submissions or internal capital models, ensure PD methods, PIT/TTC treatment and mapping logic are transparent, auditable and evidence-based.

Limitations & best practices

Limitations: PD estimates are model-dependent and sensitive to data, assumptions and economic shifts. Market-implied PDs reflect risk premia and liquidity; structural models rely on strong assumptions. Rare default events create statistical uncertainty; account for model risk in capital and governance.

Best practices:

Triangulate using statistical, structural and market approaches where feasible.
Separate discrimination testing from calibration testing.
Clearly document PIT vs TTC status and conversion methods.
Handle censoring and survivor bias with survival analysis and vintage tracking.
Maintain regular backtesting, recalibration schedules and independent validation.
Report sensitivity ranges for recovery and macro assumptions used in market-implied PDs.

FAQ

How is PD different from the observed default rate?

PD is an estimated probability for an individual obligor over a horizon; the observed default rate is a historical proportion of defaults in a cohort. PDs can be PIT or TTC; observed default rates are retrospective.

What is the typical horizon for PD?

One year is standard for regulatory and many commercial uses, but multi-year PDs are used for portfolio planning or lifetime expected loss (e.g., IFRS 9).

How often should PDs be recalibrated?

PIT PDs often need quarterly recalibration or when macro conditions change; TTC PDs recalibrated less frequently (annually). Both require ongoing monitoring.

Which method is best — statistical, structural or market-implied?

It depends on data and use-case. Statistical methods are versatile and common; structural models suit listed corporates; market-implied PDs are timely but noisier.

How do I convert model scores to PDs?

Use logistic mapping, Platt scaling, isotonic regression or empirical binning with observed default rates. Validate out-of-sample.

How sensitive are market-implied PDs to recovery assumptions?

Very sensitive — small changes in assumed recovery materially change implied PD. Report sensitivity ranges.