RESEARCH DOSSIER
Trade Credit Insurance PD LGD EAD and Exposure Measurement
Abstract. Trade credit insurance transfers a contractually defined part of the loss arising from the nonpayment of commercial obligations. Its actuarial measurement requires the joint reconstruction of buyer default, the economic loss remaining after recovery, and the eligible amount exposed when the relevant event occurs. This monograph develops a comparative synthesis of probability of default (PD), loss given default (LGD), and exposure at default (EAD), distinguishing bank credit-risk measures from the economic loss of the supplier and the contractually determined cash flows of the insurer. It examines the historical and institutional origins of these parameters, the identification of default cohorts, duration and competing-risk estimation, the valuation of recoveries, revolving invoice exposures, policy-level coverage constraints, portfolio dependence, the measurement of aggregate capacity and turnover, and underwriting and reserving feedback. Dated Basel Framework chapters, British and Indian insurance policy specimens, United States liquidation law and export-credit guidance, European insurance prudential rules, accounting guidance, member-reported market evidence and scholarly studies provide the documentary basis. Worked examples isolate the effects of coverage attachment, partial insurance, recovery timing, limit use, and dependence between event severity and exposure. The analysis establishes that the familiar product of PD, LGD, and EAD is exact under a specified conditional factorization and can become misleading when separately averaged quantities are multiplied without regard to their observation units or dependence. The research integrates existing actuarial, legal and economic methods; it makes no claim of a newly discovered empirical parameter or a new named theory.
Keywords: Trade credit insurance; probability of default; loss given default; exposure at default; accounts receivable; buyer risk; insured debt; credit limits; default hazard; protracted default; recovery; subrogation; credit concentration; economic loss; expected loss; capital; model validation.
Introduction
Credit sales create an asset for the supplier and a payment obligation for the buyer. The asset can disappear economically long before it is removed from the accounting records. A missed due date, an inability to refinance, an insolvency filing, the rejection of a claim and the eventual distribution from an insolvency estate refer to different moments in the life of the obligation. Insurance adds a further sequence of contractual decisions, including the granting of cover, attachment of an individual transaction, the application of a limit, notification and proof of loss. A rigorous measurement of insured credit risk has to represent both sequences.
The research object is short-term commercial receivables insurance, with official export-credit cover employed as a documented contractual comparator. Pre-delivery, political, medium- and long-term credit and surety exposures enter only when they clarify the boundary of post-delivery credit risk. The inquiry addresses three connected questions: how the probability of buyer default can be estimated under a well-defined observation protocol; how recovery rights, cash flows and settlement affect severity; and how the stock of receivables actually exposed at the relevant event can be reconstructed and converted into an insurer's obligation.
The method combines dated documentary analysis, examination of a published policy specimen, a targeted literature review, and reproducible analytical examples. The international comparison concerns the Basel banking framework, a European insurance capital rule, British official and Indian private credit-insurance specimens, United States liquidation law and export-credit guidance, international accounting scope and member-reported cross-border trade insurance evidence. It is deliberately a comparative study of identifiable instruments rather than a comprehensive survey of every national insolvency statute or private insurance form. No individual executed insurance contract, proprietary claims panel, buyer credit bureau or complete recovery cohort was available. Consequently every numerical example developed here is hypothetical; reported market numbers and empirical results are attributed to their original observation populations. Regulatory instruments are identified by the examined version or update date. The research date is November 18, 2026.
A central methodological distinction concerns the question being answered. The supplier asks what present value of the debt will remain unpaid; the insurer asks how much it will disburse and recover under a valid policy; the capital model asks how adverse outcomes aggregate across the book; the accountant applies a defined recognition and measurement framework to a specified reporting entity. These quantities can share data and methods while having distinct contractual and temporal boundaries.
Conceptual and Historical Foundations
From trade obligations to risk parameters
The commercial practice of selling goods on account creates credit through the operating cycle rather than through the direct disbursement of a bank loan. An invoice therefore simultaneously documents a sale and a future cash receipt. Its maturity is ordinarily linked to commercial payment terms. Repeated shipments generate overlapping debt vintages, and repayment of one invoice can coincide with attachment of later transactions. The underlying exposure is a moving stock, not the total value of sales made during a year.
Private trade credit insurance became an increasingly important institutional component of this form of finance. Van der Veer (2015), in his study of export credit insurance, situates its modern European development in the expansion of private short-term cover and the debate surrounding the 2008–2009 collapse in trade. The historical relevance of that episode is the mutual dependence of the credit and insurance markets: deterioration in payment outcomes can affect the terms and availability of protection for subsequent shipments.
The conceptual vocabulary PD, LGD and EAD has a mature regulatory expression in banking. The 2023-effective Basel IRB risk components specify a one-year PD associated with a corporate borrower's internal grade and distinguish risk components, including the treatment of certain undrawn commitments. The accompanying minimum IRB requirements describe how bank risk estimates relate to the reference definition of default and economic loss. This is a useful common language, but its jurisdiction, observation unit and purpose must be retained when using it as a comparator.
For this study, the probability of default is the conditional probability that an obligor enters a specified default state during an identified future interval, given its information and eligibility at the observation origin. Loss given default is the fraction of a defined debt exposure that is lost economically conditional on that event, or, if explicitly described as insurer LGD, an insurer's specified net cash loss divided by a corresponding insured exposure. Exposure at default is the relevant outstanding monetary exposure immediately before or at the defined default event according to an established valuation rule. These are operational definitions used for analysis, not proposed replacements for regulatory or policy definitions.
Each definition needs a time horizon, population and information set. A 12-month buyer PD and a 90-day invoice nonpayment probability answer different questions. An ultimate LGD calculated from a closed recovery file and an incurred claim-to-exposure ratio calculated after three months of development refer to different stages. EAD measured as gross buyer debt, eligible insured receivables, and the insurer's remaining maximum indemnity are different monetary variables.
The contract as a measurement boundary
An indemnity arrangement ties the insurer's promise to a loss of the insured creditor under specified conditions. An accessory surety arrangement supports the principal's obligation to the obligee. Independent guarantees and documentary standby instruments can involve an obligation to pay against specified demands or documents. Financial guarantee arrangements can support monetary obligations under their own defined terms. These relationships may share a defaulting obligor, but their claims cannot be measured by applying a single assumed payment trigger to all instruments.
The accounting distinction is equally important. The IFRS Interpretations Committee's March 2025 guarantee analysis explains that scope depends on the actual terms and that a financial guarantee contract is generally addressed by IFRS 9, with a limited election under specified prior accounting conditions. Other issued guarantees require further scope analysis. The official IFRS 17 overview identifies a separate measurement and reporting standard for insurance contracts within its scope. Thus the issuer's industry label alone cannot establish the accounting classification.
The UK Export Finance (UKEF) multiple-contract policy specimen supplies an observable example of the contract architecture. The parties are the exporter and the Secretary of State acting through the Export Credits Guarantee Department, the buyer is identified in the schedule, and both insured percentage and maximum liability are contractual quantities. Clause 4 addresses contract eligibility and attachment conditions; clause 3 defines the credit-period loss and its proof conditions. An exporter may have a commercial debt even where those coverage conditions are unmet.
A lender that receives an indemnity as loss payee is a further distinct participant. The specimen clauses 9 and 10 address payment direction and insurer subrogation; its limitations include restrictions on transfers of rights. Measurement should therefore separate identity of the buyer, insured seller, payee, creditor holding the receivable, and ultimate beneficiary of recoveries. It should also distinguish a payment direction from an effective assignment, which depends on the actual agreement and applicable law. This monograph uses the specimen to demonstrate these distinctions without making a universal claim about assignment or creditor rank.
Expected loss and its conditioning
Define a buyer default event D by horizon H, a nonnegative gross outstanding exposure X at the event, and an economic fractional loss G conditional on default. Set the loss outside D to zero. Then the expected loss of the measured debt is
EL = P(D) E[XG | D].
For a positive conditional mean exposure, define EAD = E[X | D] and exposure-weighted LGD = E[XG | D] / E[X | D]. This gives EL = PD x EAD x LGD as an exact identity. The LGD here is weighted by default exposure; it is not necessarily the arithmetic mean severity across defaults.
Alternatively, with ordinary default-conditional sample means, E[XG | D] = E[X | D] E[G | D] + Cov(X,G | D). A positive conditional covariance makes multiplication of the separate means understate the true joint expectation. If exposure is fixed and loss is measured on that same exposure, the covariance vanishes.
The banking IRB expected-loss treatment establishes a regulatory PD-LGD-EAD structure for relevant nondefaulted exposures, while specifying different treatment after default. Its parameters and requirements have a defined banking purpose. The exact conditional identity above explains the mathematical connection independently of the regulatory prescription.
Insurance changes the loss variable. Let C be a valid covered-event indicator, I the resulting indemnity, K the allocated claim expenses, and R the insurer's subsequent recoveries. With a valuation-date discount factor v(t), the present value of insurer loss is the discounted indemnity plus expenses less discounted recoveries, conditional on C. The relevant expected loss is E[1(C)L], where L denotes that insurer cash-flow value. If a covered event can arise from a political interdiction rather than the buyer's own financial failure, C need not be a subset of D. A model must partition causes or explicitly model overlapping triggers to avoid double counting.
PD Estimation
Definitions and clocks
Default estimation begins with a dated event dictionary. In the banking comparator, the Basel Committee (2022) defines default using indicators of an inability to repay or material delinquency beyond a prescribed threshold; its default and estimation provisions require consistent recording and adjustment of conflicting external data. An insurer may observe the same adverse buyer conditions, but policy events and waiting periods govern the claim rather than the bank's reference state.
UKEF's policy administration guidance illustrates the temporal separation. It calls for early notification, identifies overdue-payment reporting and ordinarily permits nonpayment claims after six months of continued nonpayment, subject to policy wording. In the specimen, clause 3 provides for shortening the waiting period on particular evidence and addresses disputed debt. A claim reporting date can consequently be later than the earliest evidence of credit deterioration; a claim payment can be later still.
A database requires at least the invoice date, contractual maturity date, original and amended payment terms, first delinquency observation, selected credit-default date, covered event date, notification date, coverage decision date, claim settlement date and recovery cash-flow dates. Amendments and retrospective records should retain both effective dates and entry timestamps. The event used for PD must be specified without using facts that would have been unavailable at prediction time.
Suppose a buyer misses a payment, pays in part two months later, and enters insolvency another month afterward. A delinquency model, a buyer insolvency model and an insurance claim model may create three distinct targets. A contractual payment dispute could withhold claim eligibility even if the buyer's credit state is deteriorating. Conversely a solvent buyer could face transfer restrictions. The dataset should preserve these paths rather than force them into an undifferentiated binary claim flag.
Cohorts and estimation
For a buyer-level annual PD, select obligors alive and nondefaulted at the prediction date, assign each a stable identifier, and observe the first qualifying default during the next 12 months. If observation is complete, an unweighted cohort estimator is d/n, with d defaulting obligors and n eligible obligors. The Basel Committee's long-run PD provisions use count-weighted obligor observations for specified bank exposures, providing a useful example of why the unit of analysis matters.
Consider 1,000 hypothetical eligible buyers with 20 first defaults observed over a complete year. The cohort PD is 2%. If the 20 defaults happen to represent 6% of original exposure, the exposure-weighted default fraction is 6%, not a competing estimate of the same count-weighted probability. The latter is useful for concentration analysis, provided that the weighting rule is disclosed.
If buyers enter at different dates, or if the portfolio contains many short-lived relationships, use a duration model. For interval k let h_k denote the conditional event probability given survival to its beginning. Survival through K intervals is the product of (1-h_k); cumulative PD is one minus that product. The model can incorporate changing rating, sector, geography, arrears and macroeconomic state, as long as predictors are known at the relevant observation time.
With a constant annual PD p, an illustrative equivalent duration under a constant hazard is p(t) = 1 - (1-p)^t, with t measured in years. For p = 2%, the 90-day approximation using a quarter-year is 0.5038%. Four independent identically sized quarterly conditional probabilities of this kind compound to the annual value, but four invoices from the same buyer are not four independent annual default events. The conversion is invalid if hazards are materially seasonal, risk grades change, or the contract has an exposure window inconsistent with the chosen interval.
When withdrawals occur before the end of observation, survival methods can handle right censoring under their relevant assumptions. Coverage cancellation, however, may be informative: an insurer may withdraw precisely because risk is worsening. Treating those buyers as ordinary unrelated censoring can understate deterioration in the original risk population. A separate follow-up stream for withdrawn, declined and exhausted-limit buyers supports assessment of this selection effect.
Predictors and calibration
Financial statement variables capture capacity and liquidity, while invoice histories capture payment behavior in a specific commercial network. An underwriting model may include leverage, cash coverage, profitability, age of the accounts, data vintage, overdue share, payment volatility and deterioration relative to a comparable industry. These features require explicit missingness indicators and time stamps. The absence of updated accounts may reflect delay or a lack of data rather than any uniform signal of insolvency.
A transparent logistic model gives a probability for a fixed horizon; a discrete-time hazard model gives probabilities conditional on survival in successive intervals. A model of quarterly scores trained on annual default outcomes must account for overlapping windows and repeated buyers, or standard errors and validation results can be misleading. Holdouts should be chronological and group-sensitive. A random split of invoice rows from the same buyer into both training and test data leaks relationship information.
Calibration and ranking are distinct objectives. A ranking statistic can improve while predicted probabilities remain systematically low. For a validation cohort of size n, the Brier score is the mean of (p_i-y_i)^2, where y_i is the observed binary event. Comparison of predicted and realized default frequencies by grade and time period complements ranking, with uncertainty represented by confidence intervals or suitable resampling. A model fitted to unusually stable years should be tested on adverse periods before long-run assumptions are accepted.
Point-in-time probabilities reflect conditions known at the forecast date. A long-run estimate averages experience over a suitable distribution of economic states. The latter is not obtained simply by suppressing every macroeconomic variable from a score. The choice needs to follow use: responsive underwriting and short-term forecasting can benefit from the former, while capital planning and comparisons across cycles may require long-run and stressed estimates. For banks, the Basel framework's long-run requirements are specifically defined; they are not an insurance-wide calibration mandate.
Competing causes and operational selection
For insurance, insolvency, prolonged nonpayment, and political interruption can be modeled as separate causes of a covered event where the contract supports these categories. If h_j(k) denotes a mutually exclusive cause-specific hazard in interval k, the all-cause hazard is the sum over causes; cumulative incidence for cause j sums survival to each interval multiplied by its hazard. A legal insolvency discovered after a commercial nonpayment must receive consistent event-priority treatment if the model uses exclusive first causes.
The insurer controls part of the observed population by approving risks, granting limits and changing cover. Yang, Bakshi and Chen (2021), in Trade Credit Insurance: Operational Value and Contract Choice, identify monitoring and cash-flow smoothing as different roles and analyze contract choice within a strategic model. The implication for estimation is that a limit is an intervention as well as a descriptive field. Buyers with reduced limits may display higher subsequent default rates because of risk deterioration; the association alone does not establish that the limit reduction caused the deterioration.
The calibration data should distinguish applicants, approved buyers, active insured invoices and claims. When rejected applicants lack follow-up observations, estimation of population-wide default remains partially unidentified. Report the model as conditional on the observed approved population unless external follow-up or a justified selection adjustment supports broader inference. The research does not generate a calibrated PD for an actual insurer because no such panel was obtained.
LGD Recovery
Cash-flow reconstruction
Economic severity is a function of recovery amount, recovery timing, associated cost and the definition of the initial exposure. The Basel Committee's economic loss definition explicitly treats material collection costs and discount effects as relevant for bank LGD. An insurer's recovery process requires analogous valuation decisions, but the insured share, claim terms, legal standing and insurer-specific expenses determine which cash flows enter its own model.
For gross buyer debt X at default, let r_j be recoveries at delays t_j and k_j relevant collection costs. With a fixed illustrative annual discount rate a, the discounted economic loss at default is
L_debt = X - sum[r_j/(1+a)^t_j] + sum[k_j/(1+a)^t_j].
Debt LGD equals L_debt/X when X is positive. A zero-exposure event contributes to the default count but cannot have an ordinary ratio defined by that denominator. Separate genuine zero exposures from missing exposures and specify treatment of recoveries exceeding the originally recorded debt or costs exceeding it. An unconstrained economic ratio may fall outside zero to one; contractual indemnity calculations may still be bounded differently.
A hypothetical receivable of USD 100,000 at default yields USD 30,000 after one year and USD 10,000 after two years, with USD 5,000 collection costs after one year. At 5% the discounted net recovery is 25,000/1.05 + 10,000/1.05^2 = USD 32,879.82. Economic loss is USD 67,120.18, or 67.12%, whereas undiscounted net loss is USD 65,000. The example demonstrates that a nominal recovery ratio alone omits the timing of collection.
A creditor's protection from the insurer is an additional cash inflow for the creditor and an outflow for the insurer. It cannot be subtracted from both parties' losses when evaluating the underlying buyer's debt severity. The consolidated perspective has to eliminate internal transfers to avoid treating insurance as creating recovery from the buyer.
Contractual transformation and timing
Under the specific UKEF specimen, clause 3.1 ties nonpayment indemnity to the outstanding insured debt after the waiting period, subject to other policy provisions. Clause 13 distinguishes creditor recoveries before claim payment from the allocation of recoveries after payment. The specimen recovery provisions give a direct example of how the same underlying debt produces different insurer cash-flow timing according to the stage when repayment occurs.
Consider a simplified hypothetical private policy, not a quotation of UKEF terms. The buyer owes USD 100,000 at default. Eligible debt is USD 80,000, a preclaim eligible collection of USD 10,000 reduces it, and the insurance percentage is 90%. Assume a remaining policy cap of USD 60,000 measured in indemnity units, no deductible and no other applicable exclusion. The initial claim is min(60,000, 0.9 x 70,000) = USD 60,000. Suppose a further USD 20,000 is recovered one year after the claim, and the illustrative contract assigns 90% of that recovery to the insurer. If the claim and USD 2,000 expenses occur at the valuation origin and discounting is 5%, insurer economic loss is 60,000 + 2,000 - 18,000/1.05 = USD 44,857.14. Had the cap been absent, the initial claim would be USD 63,000. The difference is a policy effect, not a change in the buyer's PD.
The UKEF exporter guide describes a typical maximum insured percentage of 95% for its product. This is evidence about an official export-credit provider, not a globally applicable private-market percentage. The specimen's schedule allows the insured percentage and maximum liability to be specified. Consequently a model should store these as effective-dated variables. It should also preserve whether a maximum is defined on eligible debt, indemnity payable or aggregate policy payments; the same numeric value in different units yields different claim arithmetic.
Claims development also includes partial payments, additional approved amounts, reopened cases and recovery expenses. Reserving at a reporting date requires expected remaining future amounts, rather than simply accumulating paid claims. Gross incurred claims, gross ultimate claims, net of buyer recovery claims, and net of reinsurance results should be reconciled at separate stages.
Recovery heterogeneity and legal mechanisms
Recovery can arise through payment plans, distributions in formal insolvency, enforcement of security, retention of title, settlements, or payments from another legally obligated party. Their amounts and dates depend on the jurisdiction and the enforceability of the relevant rights. Insolvency distributions require information on priority, estate value, competing creditors and procedural expense. A documentary payment obligation of a bank is a distinct source of repayment; the existence of a letter of credit in a transaction is insufficient without determining whether its terms and presentations support collection.
Clauses 6, 10 and 12–14 of the UKEF contract wording address mitigation, subrogation, allocation, recoveries and currency. The existence of such conditions demonstrates why recovery attribution cannot be inferred solely from an insolvency label. Actual recoveries should retain the recipient, original source, currency, payment date, linked claim, cost allocation, policy share and legal basis. A later transfer between insurer and exporter should be tagged as a distribution of a recovery, not an additional recovery from the buyer.
Severity estimation can use a two-part structure when the conditional distribution includes a mass at zero recovery or full recovery: estimate the probability of the relevant extreme, followed by a model of intermediate amounts. Fractional regression is appropriate only when the severity definition is actually bounded. Recoveries extending over multiple years require development methods or cash-flow prediction rather than restricting calibration to quickly closed claims.
Suppose 50 default cases are observed and only 30 are closed at the review date. Computing an ultimate recovery mean from the closed cases implicitly assumes that closure is uninformative about severity and timing. In practice short and uncomplicated cases can close earlier, and litigation-rich cases may remain open for years. Use actual observed cash flows through the valuation date, project remaining amounts under explicit assumptions and revisit earlier estimates as mature cohorts develop. The resulting uncertainty should be reported rather than concealed by dropping the 20 open cases.
The legal route of recovery is a separate conditioning variable. In the United States, Chapter 7 liquidation under 11 U.S.C. § 726 distinguishes statutory priority claims, permitted categories of allowed unsecured claims, and other classes in the distribution sequence. The amount of receivables recognized by the insurer therefore cannot establish the creditor's expected distribution. Claim allowance, legal priority, available estate and timing require separate evidence. This comparison is limited to the cited liquidation provision; reorganizations, security enforcement and other proceedings require their own legal analysis.
The Indian shipment-policy specimen issued by Generali Central Insurance Company Limited (n.d.) provides a complementary contractual perspective. Its subrogation and recovery conditions distinguish pursuing recovery from allocating the proceeds after indemnification. Under D.18 the insurer's participation in post-payment recoveries continues until its payment and recovery costs have been reimbursed, with subsequent amounts going to the insured. This is a contractual allocation rule rather than a universal recovery rate. Where a creditor collects an amount from an estate, the economic recovery to the creditor and the insurer's reimbursement should be reconciled through this rule rather than independently entered as if both parties had received the same money.
Dependence and economic downturns
Altman, Brady, Resti and Sironi (2005), in their published default and recovery study, report a relationship between corporate bond default rates and recoveries using 1982–2002 data. That result motivates a stress hypothesis for trade credit, where default clustering could coincide with lower secondary realization values or business asset proceeds. Bond recovery magnitudes are not transferable to commercial invoices without evidence of comparability.
For illustration, a portfolio with 1% PD and 40% conditional LGD in a normal state, versus 5% PD and 70% LGD in distress, has state losses of 0.4% and 3.5% of a fixed exposure. With respective state probabilities of 80% and 20%, the unconditional expected loss rate is 1.02%. Averaging PD and ordinary state LGD separately gives (1.8%)(46%) = 0.828%. The distinction is the dependence between default incidence and severity. The unconditional default-weighted severity is 1.02%/1.8% = 56.67%. This hypothetical example is mathematical, not an estimate of a real portfolio.
Recovery resources can themselves become constrained under concentrated stress. A larger number of simultaneously open cases may extend collection times and raise expense. This mechanism is plausible independently of changes in estate value; it should be tested by controlling for case type and workload rather than attributing all deterioration to macroeconomic conditions.
EAD Measurement
Stocks, flows and insurance capacity
Trade-credit portfolios revolve through shipments, invoicing, adjustment and collection. At any valuation instant, gross receivables equal opening unpaid debt plus new invoices and authorized adjustments less payments, credit notes and writeoffs applicable to the chosen debt definition. A reconciliation must establish which removals are accounting events and which extinguish the legal claim. A writeoff may change recognized carrying amount while recovery rights continue.
Let X(t) denote unpaid debt at time t. Let A(t) be the sum of individual eligible receivables at that time under the relevant policy version. Then A(t) can be less than X(t), for instance because some shipments did not attach or some receivables fall outside the contractual scope. A buyer limit B(t) and an insurance percentage q(t) may further restrict the indemnity. A policy-level remaining liability M(t) separately governs aggregate exhaustion where such a clause exists.
For a hypothetical simple limit in debt units, insured debt before recovery is min(A(t), B(t)). If the contract applies a fixed percentage q to remaining covered loss and has a payout-unit cap M, a simplified indemnity is min(M, q x max(0, min(A,B)-R_pre)), where R_pre is the applicable preclaim recovery. This is an illustrative ordering only. In a real case the waterfall must follow the contract's rules on how payments and limits attach; a deductible or different cap position changes the expression.
The UKEF maximum-liability clauses show that the specimen expressly caps aggregate payable amounts. Its schedule also contains an alternative maximum tied to the insured percentage of a permitted credit limit under a specific endorsement. It is therefore inappropriate to assume that every figure called a credit limit equals the payout cap. Both the source of the variable and the unit of the cap require verification.
An insurer may report approved limits, utilized limits, insured shipments, turnover, gross open balances or maximum payable amounts. Each supports a different question. Approved capacity is a potential stock; turnover measures transactions over an interval; utilization relates covered outstandings to available capacity at a date. EAD is conditional on the relevant adverse event and cannot be read directly from any single aggregate.
Invoice ledger and a worked portfolio
The following amounts, in USD thousands, describe an illustrative default-date buyer snapshot. The limit is an eligible-debt-unit cap; assumed indemnity rate is 90%; no recovery or deductible has yet been applied.
| Item | Gross receivable | Eligible receivable |
|---|---|---|
| Invoice A | 40 | 40 |
| Invoice B | 35 | 35 |
| Invoice C | 25 | 0 |
| Total | 100 | 75 |
With a debt-unit limit of 60, gross buyer EAD is 100, eligible exposure is 75, capped debt exposure is 60, and initial maximum payout from this snapshot is 54, subject to any additional policy cap. Multiplying the 100 gross debt by the policy's 90% participation gives 90 and overstates this hypothetical maximum by 36. Multiplying the approved limit by PD to approximate expected loss would additionally omit severity and any state-dependent utilization.
The stock is also sensitive to time. A hypothetical supplier selling USD 1.2 million uniformly over a 360-day convention with 60-day payment terms and timely collection has steady-state receivables of approximately USD 200,000, assuming no seasonality, no cash sales, and no initial or terminal adjustments. Increasing average collection duration to 90 days under identical assumptions produces USD 300,000. The annual sales flow is unchanged while the outstanding debt stock increases 50%. A buyer's actual default exposure depends on shipment timing, payment behavior and coverage rather than on this steady-state approximation alone.
The EAD forecast should account for future invoices only where the chosen exposure origin and contractual scope permit them. Existing uncovered purchase orders are not automatically insured exposures. A limit granted today may be withdrawn, reduced or exhausted before a later shipment; the treatment of prior attached cover depends on the policy. The insurer can still face claim obligations for legitimately attached transactions after a limit has been changed.
The Generali Central specimen also supplies a concrete example of the temporal distinction between attachment and default. Its claim calculation conditions (General Conditions E.14) use the buyer limit applicable at shipment, track losses in relation to an aggregate deductible, restrict recoverable losses by additional limits and specify currency conversion dates. Clause D.8 additionally distinguishes individual buyer and group limits, while D.12 restricts subsequent shipments after specified adverse events. An EAD data model that overwrites the shipment-date authorization with a lower limit observed at default can misstate contractual exposure. A policy-year ledger should preserve the histories of each limit and retention account alongside the buyer-level monetary snapshot.
For illustration, consider an eligible shipment of 100, a limit of 100 at shipment, a current limit of 60 after an adverse monitoring event, and a subsequent default with no recovery. Under a hypothetical contract that preserves the original attachment and applies 90% indemnity, the amount before other applicable caps is 90. A calculation using only the current limit produces 54. The difference of 36 is a data-history error under the stipulated terms, not a change in the buyer's realized economic severity. Actual limits and coverage depend on the complete contract, which may include other conditions not represented in this example.
Exposure growth before default
Buyers in distress can accumulate unpaid invoices even as suppliers try to stop shipments. If the insurer's information is delayed, reported EAD may underestimate the exposure that existed at the date of the credit event. The model needs to distinguish additions before the actual event from later invoices and post-event accounting changes.
A banking EAD model can estimate future utilization of undrawn credit lines; the Basel Committee's IRB exposure components prescribe rules for relevant commitments. A rolling trade-credit relationship is economically analogous in that a stock may rise before default, but an approved insurance limit is not itself an obligation to lend. Bank regulatory conversion factors cannot be assigned mechanically to an insurance limit.
A conditional EAD model can predict the outstanding stock at an event from opening debt, expected eligible shipments, subsequent payments and the time-to-default distribution. The forecast should enforce legal and accounting reconciliation rather than relying on a single utilization ratio. A negative stock after payment is handled as a credit balance according to the ledger convention; it is not an admissible negative insured debt. Incomplete currency data, duplicate invoices and unrecognized credit notes should be resolved or quantified before portfolio aggregation.
A simple stress example uses a current eligible exposure of USD 100,000, conditional future net accumulation of USD 30,000 before default, and a debt-unit limit of USD 120,000. The conditional eligible stock is USD 130,000 and the limit-adjusted stock is USD 120,000. If late payments or ineligible later shipments invalidate the accumulation assumption, this projection changes. Forecast error is thus a combined function of credit behavior, operational volume, reporting lags and coverage attachment.
A second public export-credit comparator is the Export-Import Bank of the United States (EXIM). Its Multi-Buyer Standard Insurance guidance distinguishes coverage percentages for ordinary commercial and political risks from specified sovereign and bulk agricultural exposures. It also describes buyer-specific credit authorization mechanisms. Product design can consequently create variations in the insurer's expected indemnity even when two exposures share the same estimated buyer PD and economic LGD. This product guidance supplies evidence of institutional design, not a substitute for applicable policy wording or a frequency sample.
Market data and international comparison
The Berne Union's State of the Industry 2025 reports USD 3.345 trillion of short-term trade insured during 2025 and approximately USD 2.58 trillion of outstanding short-term revolving limits at year end. It also reports USD 11.107 billion of total claims paid across business lines. The report identifies foreign-exchange effects in the year-end limits comparison. These are member-level reporting aggregates of different types, not matched default-cohort observations.
Dividing total claims paid across business lines by short-term insured trade gives about 0.332%, a mechanically calculable number with incompatible numerator and denominator. It is not a valid short-term claims rate, much less an annual buyer PD. Even short-term claims paid divided by short-term turnover would remain a payment-flow intensity, unless matched by event, contract scope, reporting population and exposure development. Reporting an exposure-based ratio requires a consistent average stock or other explicitly defined exposure base.
Koehler's (2023) post-pandemic claims analysis shows that claims trajectories differed between product lines and discusses lags following shocks. Market evidence must therefore be interpreted by product, loss occurrence and payment timing. A country comparison also depends on the collection process, commercial payment terms, composition of insured buyers and the currencies in which figures are expressed. Changes in the mix can change aggregate metrics even when each homogeneous segment's risks remain constant.
Evidence Verification and Model Sensitivity
A joint scenario calculation
Consider an illustrative model with two equally likely economic states. Conditional on the normal state the buyer PD is 1%, EAD upon default is USD 100,000 and loss severity upon default is 40%. Conditional on the stress state the buyer PD is 5%, EAD is USD 140,000 and severity is 75%. Assume within-state EAD and LGD are fixed for this example, the same horizon applies to all quantities, every event has the same debt scope, and discounting has already been incorporated in LGD.
| State | Weight | PD | EAD | LGD | Weighted expected loss |
|---|---|---|---|---|---|
| Normal | 0.50 | 0.01 | 100,000 | 0.40 | 200 |
| Stress | 0.50 | 0.05 | 140,000 | 0.75 | 2,625 |
| Total | 1.00 | 2,825 |
Independent multiplication of the state-weighted averages gives 0.03 x 120,000 x 0.575 = USD 2,070. The joint expectation is USD 2,825, so the naive result is lower by USD 755, or 26.73% of the joint expectation. The difference arises because default selects adverse states more frequently and the exposure and loss severity are also larger there.
To recover the exact factorization, unconditional PD is 3%. The expected default-exposure amount before multiplying by severity is 0.5 x 0.01 x 100,000 + 0.5 x 0.05 x 140,000 = USD 4,000. Dividing by PD produces default-conditional EAD of USD 133,333.33. The exposure-weighted conditional LGD is 2,825/4,000 = 70.625%. These choices give the joint expected loss. The mathematical identity does not require independent parameters; it requires correctly weighted conditional quantities.
Portfolio concentration and shared buyers
A buyer can owe money to many insured suppliers simultaneously. With hypothetical debt losses of USD 60,000 and USD 40,000 on two policies responding to the same buyer bankruptcy, the buyer-event aggregate is USD 100,000 before considering policy terms. If the probability of this single event is 2%, the aggregate expected loss is USD 2,000. Treating the two policies as independent events does not change their sum of means, but it materially changes the distribution.
For the shared-event model the probability of both payouts is 2%. Under a false independent-event assumption it is 0.02 x 0.02 = 0.04%. At a 99% quantile with the conventional smallest-loss-at-cumulative-probability rule, the true shared-event loss is USD 100,000, whereas the false independent distribution gives USD 60,000. The discrepancy shows why buyer entity resolution is a capital concern as well as a reporting concern.
Group identifiers are equally important when legally distinct buyers depend on the same parent, funding channel or supply chain. They should not automatically be merged into a single legal debtor, because policy and recovery rights may differ. Preserve legal identities, link common risk groups, and model conditional dependence separately. Exposure aggregation across groups, sectors, countries, currencies and policyholders should use the same valuation timestamp.
Insurance capital and accounting boundaries
The March 2024 version displayed by EIOPA of Solvency II Article 134 defines a credit-and-suretyship shock sub-module combining a large-credit-default and recession component through square-root aggregation. It specifies an instantaneous default of the two largest relevant exposures, a gross loss assumption of 10% of sum insured and selection by net loss after recoverable amounts; the recession component is defined with reference to the following 12 months' earned premiums.
This provision exemplifies a prudential stress prescription for a particular European standard-formula component, not a general-purpose estimate of observed buyer LGD. It is a regulatory shock to own funds, with reinsurance treatment, and the displayed article is only one part of the insurance capital framework. A full standard-formula calculation would require the other applicable modules and rules, which are outside this dossier. The numerical examples here do not claim to calculate a firm's SCR or compliance position.
The supplier's accounting problem is different. The HMRC IFRS 9 impairment guidance describes expected cash shortfalls and a simplified approach for eligible receivables that uses lifetime ECL, including time value and reasonable forward-looking information. A 12-month expected credit loss concerns lifetime losses from defaults possible in the next 12 months; it is not simply the amount of unpaid bills expected to appear within one year. The guidance is an explanatory source, not a substitute for the full current accounting standard. The treatment of insurance proceeds or integral credit enhancement is contract- and standard-specific; this research does not assume automatic netting of an independently held insurance policy against a receivable's ECL.
For an insurer, the scope analysis stated in the IFRIC guarantee discussion is relevant to the classification of particular issued obligations. Once the accounting scope is determined, the applicable measurement standard governs. An actuarial estimate of expected discounted claim cash flows is an input into broader reporting and management processes; it does not by itself establish insurance contract liabilities, revenue, risk adjustment or contractual service margin.
A recent empirical contribution by Yoo, Tseung, and Fung (2025) links claim probabilities to the network of trade connections while explicitly modeling reporting delay. Their trade credit network study uses 294,272 connections in a proprietary insurer dataset, with 2020 observations held out from the preceding five years. In the reported data approximately 2.5% of connections generated a claim and 5.66% of policies reported at least one claim, observed through June 30, 2023. These frequencies differ because their denominators differ. Neither estimates the annual PD of unique legal buyers. The study's treatment of incomplete claim reporting and chronological evaluation is directly relevant to observation maturity and dependence in the present monograph. Its preprint supplies published empirical evidence of a specific portfolio; the proprietary data were unavailable for replication, and the study is not treated as independent validation of the illustrative calculations here.
Data quality and verification protocol
For each active buyer, preserve a valuation-date exposure snapshot reconciled to individual invoice records; for each default, preserve event-definition evidence and the exact snapshot generated as of the event. Maintain a separate claims register with policy and endorsement identifiers, loss dates, approved and paid amounts, cap utilization, claim reasons, expenses and recoveries. A policy excluded for a material condition must not be represented as a zero-severity economic default without a coverage indicator.
Exposure accuracy is tested by comparing reconstructed outstandings to contemporaneous ledger snapshots and tracing adjustments to their source. For a matured default cohort, test whether the exposure model systematically overstates or understates actual event-date outstandings by buyer segment and horizon. Severity accuracy is tested on both nominal and discounted outcomes, with open-case projections evaluated against later development. PD accuracy is assessed against the selected buyer-event definition even for buyers with no insured losses. The complete joint loss forecast must also be validated; accurate marginal models can yield inaccurate total loss if their dependencies are misspecified.
Uncertainty includes sampling variation, missing data, model form, contractual interpretation, economic scenarios and execution error. These components have different remedies. More observations can narrow sampling intervals; explicit policy versioning can reduce contractual ambiguity; alternate stress scenarios illuminate model dependence. A model run is reproducible when the underlying data cut, code version, parameter set, legal interpretation and currency convention are recoverable.
The examination of sources for this manuscript used named versions and sections of regulations and a published contract specimen. Its extracted text was accessible; the requested visual page render was unavailable, so the contract analysis relies on that text. The corporate-bond and operations-management studies are used only for conclusions supported by their accessible abstracts. The findings from Van der Veer's working paper are attributed to its own historical data and design; the underlying proprietary dataset was not obtained. Aggregate industry reports cannot identify the joint parameter distribution. Private and official policy specimens and one United States liquidation provision illustrate contractual and legal variation. The comparative sample is selective and no jurisdiction-specific recovery rate is inferred. The empirical network study is cited within its reported observation scope, without independent replication. These are limitations on empirical identification and external validity.
Integrated Scientific Synthesis and Research Agenda
From the buyer to the insurance portfolio
The preceding analysis supports a practical integration of three layers of observation. The first is the buyer's event history, used for a defined PD. The second is the obligation and its realization, including exposure as it stands at default and discounted creditor recoveries. The third is the insurance contract, which transforms eligible debt loss into a payment and subsequent recovery process. These layers share the economic fact of nonpayment while retaining distinct amounts, dates and rights.
A reliable expected-loss estimate first identifies the event and the debt stock linked to it. It then values recoveries and identifies the policy terms applicable when transactions attach and the covered event occurs. Portfolio aggregation follows legal buyer and common-risk-group mapping, after allowing for policy caps, cross-policy limits where relevant and reinsurance. Each stage adds information that cannot be inferred from the previous stage's ratio alone.
The analytical contribution is the reconciliation of the three conventional credit components with these coverage and observation boundaries. It shows why conditional weighting and contract transformation have to precede aggregation; it also shows how the same invoice ledger can generate distinct credit-stock, insured-debt and maximum-payout measures. This is a synthesis of established probability, actuarial cash-flow analysis and contract evidence rather than a proposed new named model.
Underwriting feedback and identification
The relationship between risk measurement and future exposure is endogenous. An insurer can reduce available cover after receiving adverse information, a supplier can shorten payment terms, and a buyer can seek more credit from other suppliers. Each decision changes subsequent observed exposure. A retrospective portfolio containing only retained risks will therefore describe a population shaped by previous underwriting decisions.
Van der Veer (2015) reports in DNB Working Paper 462 that, within a 1992–2006 historical bilateral dataset for a global insurer, a doubling of claims was associated in his estimated specification with approximately 11% lower share of exports insured and 4% higher premium levels; the paper also reports cross-country propagation of extreme losses. Its country-pair and time-varying national fixed effects address observed design concerns about differences between markets and broad national shocks. A single provider and historical setting, however, do not establish universal causal elasticities for every insurer or product.
The study's significance for this dossier is the possibility of a feedback from realized claims to future insured exposure and rates. Falling utilization and falling losses could jointly result from tighter underwriting rather than improved creditworthiness in the wider buyer population. To discriminate between these explanations, a future empirical design would follow all applicants or known buyers where possible, record the reasons and dates for limit changes, and distinguish changes to new shipments from treatment of preexisting cover.
Pricing, reserving and capital coordination
Expected indemnity net of attributable buyer recoveries and relevant expenses is an actuarial component of a premium calculation. It is not the entire premium. Expenses outside the claim cash-flow model, cost of capital, reinsurance cost, uncertainty and contract structure can all matter. Premium rates quoted against insured turnover require a consistent forecast of that turnover; expected loss derived from an event-date exposure stock cannot be divided by an unrelated annual stock and advertised as a turnover rate.
For already occurring defaults, a future-default probability should not be reapplied to the same event. Reserve estimates instead evaluate the outstanding indemnity and recovery cash flows arising from the occurred event, with appropriate coverage and payment uncertainty. A buyer may later return to a performing state and default again, but that requires a defined cure and new event protocol. Forecast reserves and prospective risk must be reconciled to avoid counting an existing claim both as an occurred loss and as expected future default on the same debt.
Capital management must understand both the liability-generating buyer events and the recoverability of reinsurance. Reinsurance transfers specified insurer losses under its own attachments, limits and conditions. A gross portfolio risk model and a net model should retain traceability from claim to treaty allocation and estimate the possibility that a reinsurer's financial deterioration occurs during the same unfavorable economic state.
Research agenda and empirical requirements
An internationally comparable research dataset would combine anonymized buyer identities, time-indexed ratings, exposure and contractual coverage snapshots, event and cure dates, invoice maturity and collection histories, and paid and projected claim cash flows. It would preserve rejected risks where lawful and feasible, so that underwriting selection can be studied rather than embedded invisibly in model coefficients. Privacy-preserving entity linking would be necessary for identifying shared buyer concentrations across suppliers.
The first empirical objective would establish the consistency and predictive value of alternative default labels: financial failure, specified delinquency, covered event and approved claim. The second would estimate recovery development and its dependence on jurisdiction, sector, economic state, security and case characteristics. The third would quantify how revolving exposures change before default and how limits, payment terms and coverage decisions mediate that process.
A separate methodological objective would evaluate the performance of joint forecasts versus models assembled from separately calibrated marginal components, with chronological validation and explicit exposure weighting. A causal question about the effect of limit changes on real trade requires identification beyond ordinary portfolio association, since underwriting actions respond to latent risk. Suitable data and institutional variation would determine the feasible design; no causal estimate is asserted here.
The public evidence establishes the relevance and measurable structure of such a research program, but not its empirical coefficients. That boundary preserves a useful scientific distinction between a reproducible analytical result, a published empirical result, and a proposed empirical investigation.
Conclusion
The scientific measurement of trade credit insurance begins with a time-specific commercial obligation, an identified buyer and a verified insurance contract. PD is the probability of the specified adverse event in a stated population and interval. LGD measures the economic severity conditional on that event under a defined valuation perspective. EAD measures the outstanding monetary amount exposed when the event occurs. A consistent selection and weighting of these quantities yields expected credit loss. A contract-specific cash-flow transformation yields the expected loss of the insurer.
The same adverse event can generate multiple exposures across suppliers and multiple contractual responses. It can also cause no insurance payout because the outstanding debt falls outside the insured scope, is collected during the waiting period, or does not otherwise satisfy coverage conditions. These differences must be represented explicitly in data and equations. They are part of the insured-risk measurement problem itself.
The Basel banking framework supplies a valuable comparative reference for PD, economic severity and exposure, while the UKEF specimen demonstrates the concrete significance of coverage attachment, waiting periods, caps, recovery allocation and currencies. Solvency II provides a distinct example of prescribed insurance capital stresses; accounting guidance addresses separately defined reporting obligations. Market data and published research demonstrate the importance of exposure composition, loss development and insurer feedback without supporting a universal numerical parameter for all trade credit portfolios.
The original analytical synthesis establishes the significance of default-conditioned weighting and cross-policy buyer aggregation, illustrates the economic consequences of mixing debt and indemnity units, and identifies the evidence needed to estimate the joint loss distribution. Reliable calibration for a particular book additionally requires complete, consistently labeled longitudinal exposure and recovery observations. The resulting discipline connects underwriting, actuarial valuation, claims operations, financial reporting and prudential risk in a single traceable sequence of measurements.
References
Altman, E. I., Brady, B., Resti, A., & Sironi, A. (2005). The link between default and recovery rates: Theory, empirical evidence, and implications. The Journal of Business, 78(6), 2203–2228. DOI: 10.1086/497044. Default and recovery rates
Basel Committee on Banking Supervision. (2020). Basel Framework: CRE32, IRB approach: risk components. Version effective January 1, 2023. Bank for International Settlements. IRB risk components
Basel Committee on Banking Supervision. (2020). Basel Framework: CRE35, IRB approach: treatment of expected losses and provisions. Version effective January 1, 2023. Bank for International Settlements. IRB expected losses
Basel Committee on Banking Supervision. (2022). Basel Framework: CRE36, IRB approach: minimum requirements to use IRB approach. Version effective January 1, 2023. Bank for International Settlements. IRB minimum requirements
Berne Union. (2026). State of the Industry 2025. Annual report of members' export credit and investment insurance business, reporting year 2025. State of the Industry 2025
European Commission. (2014). Commission Delegated Regulation (EU) 2015/35, Article 134, Credit and suretyship risk sub-module. EIOPA Single Rulebook, page updated March 26, 2024. Article 134: Credit and suretyship
Export-Import Bank of the United States. (n.d.). Multi-Buyer Standard Insurance. Product guidance. Accessed November 18, 2026. Multi-Buyer Standard Insurance
Generali Central Insurance Company Limited. (n.d.). Export and Commercial Perils Trade Credit Insurance (Shipments Policy): Policy Wordings. UIN IRDAN132RPTC0004V01201617. Accessed November 18, 2026. Trade credit policy wording
HM Revenue & Customs. (n.d.). CFM21890: IFRS 9: Measurement of financial assets: impairment. Corporate Finance Manual. Accessed November 18, 2026. IFRS 9 impairment guidance
IFRS Interpretations Committee. (2025, March). IFRIC Update: Guarantees issued on obligations of other entities. IFRS Foundation. March 2025 IFRIC update
International Accounting Standards Board. (2017, amended 2020). IFRS 17 Insurance Contracts. IFRS Foundation. Standard overview accessed November 18, 2026. IFRS 17 Insurance Contracts
Koehler, J. (2023, June 29). Claims: Controlling chaos, the risk versus reality. Berne Union. Claims: risk versus reality
UK Export Finance. (2022). Manage your UKEF Export insurance policy. GOV.UK. Living guidance; accessed November 18, 2026. Managing an export insurance policy
UK Export Finance. (2025). Multiple Contract Export Insurance Policy. Specimen policy wording. Export Credits Guarantee Department. Multiple Contract Export Insurance Policy
UK Export Finance. (2025). UK Export Finance's Export insurance: a guide for exporters and brokers. GOV.UK. Exporter and broker guide
United States. (n.d.). 11 U.S.C. § 726, Distribution of property of the estate. United States Code, Chapter 7. Statutory text reproduced by Cornell Law School, Legal Information Institute. Accessed November 18, 2026. 11 U.S.C. § 726
Van der Veer, K. M. (2015). Loss shocks and the quantity and price of private export credit insurance: Evidence from a global insurer. DNB Working Paper No. 462. De Nederlandsche Bank. DNB Working Paper 462
Yang, S. A., Bakshi, N., & Chen, C. J. (2021; online 2020). Trade credit insurance: Operational value and contract choice. Management Science, 67(2), 875–891. DOI: 10.1287/mnsc.2019.3571. Trade credit insurance and contract choice
Yoo, W., Tseung, S. C., & Fung, T. C. (2025). Statistical learning of trade credit insurance network data with applications to ratemaking and reserving. arXiv preprint 2509.20803, September 25, 2025. Trade credit network study