BPCE - 2018 Risk report / Pillar III

5 CREDIT RISK

Risk measurement and internal ratings

estimates are systematicallyadjusted to factor in a safety buffer for the uncertainty of the estimates. Where past internal data do not cover a full economiccycle, an additionalsafety buffer is determined in order to maintain a TTC (through the cycle) approach. For comparison purposes, risk reconciliation is carried out between internal ratings and agency ratings. Loss given default (LGD) is an economic loss measured by incorporatingall inherentfactors in a transactionas well as the costs incurred during the collection process. LGD estimation models for retail customers are applied specificallyto each network. LGD values are first estimated by product and based on whether or not any collateral has been provided. Other factors may also be considered secondarily, where they can be used to statistically distinguish betweendegreesof loss. The estimationmethodemployedis based on the observationof marginal collection rates, depending on how long the customer has been in default. The advantage of this method is that it can be directly used to estimate LGD rates applied to performing loans and ELBE (1) rates applied to loans in default. Estimates are based on internal collection histories for exposures at default over an extended period. Two safety buffers are then systematicallyadded: the first to cover estimateuncertaintiesand the second to mitigateany economic slowdown effect. Groupe BPCE uses three models to estimateEAD. The first estimatesa Credit Conversion Factor (CCF) for off-balance sheet exposures. This model is automatically applied depending on the type of product, where the off-balance sheet has a non-zero balance. A multiplying factor is specifically applied to the balance sheet for account overdrafts, where the balance sheet has a non-zero balance but the off-balancesheet has a zero balance. Furthermore,a standard EAD is applied for accounts with credit balances and no overdrafts (authorizedor unauthorized). INTERNAL RATINGS-BASED APPROACHES – NON-RETAIL CUSTOMERS Groupe BPCE has comprehensive systems for measuring non-retail customer risks, using either the Foundation IRB or Advanced IRB approachdependingon the networkand the customersegment.These systems can also be used to assess the credit quality of its loan books for better risk supervision. The rating system consists in assigning a score to each counterparty. Given the Group’s cooperative structure, a network of officers is responsible for using the uniqueness of the score to determine the customer’srating for the Group. The score assignedto a counterparty is usually suggested by a model, then adjusted and validated by Risk function experts after they perform an individual analysis. The counterparty rating models are mainly structured according to the type of counterparty(corporates,financialinstitutions,public entities, etc.) and size of the company(measuredby its annual revenue).When volumesare sufficient(SMEs, ISEs, etc.), the models rely on statistical modeling (logistic regression methods) of customer defaults,

combined with qualitative questionnaires.Otherwise, expert criteria are used, consistingof quantitativefactors (financialratios, solvency, etc.) derived from financialdata, and qualitativefactors assessingthe customer’s economic and strategic components. With respect to country risk, the system is based on sovereign ratings and country ratings that limit the ratings that can be given to non-sovereign counterparties. The Non-Retail rating scale is built using the past Standard& Poor’s ratings to ensure the direct comparabilityin terms of risks with the ratingagencies. LGD models (excludingretail customers)are predominantlyapplied by type of counterparty,type of asset, and whether or not any collateral has been provided. Similar risk categories are then defined, particularly in terms of collections, procedures and type of environment.LGD estimates are assessed on a statistical basis if the number of defaults is high enough (e.g. for the Corporate customers asset class). Past internal data on collections covering the longest possibleperiod are used. If the numberof defaultsis not high enough, external databases and benchmarks are used to determine expert rates (e.g. for banks and sovereigns).Finally,some values are based on stochastic model, for loans in collection. Downturn LGD is checked and safety buffersare added if necessary. Groupe BPCE uses three models to estimate EAD for corporates. The first estimatesa Credit ConversionFactor (CCF) for off-balancesheet exposures.This model is automaticallyapplied dependingon the type of product, where the off-balance sheet has a non-zero balance. A multiplying factor is specifically applied to the balance sheet for account overdrafts, where the balance sheet has a non-zero balance but the off-balance sheet has a zero balance. Otherwise, a standard EAD is applied for accounts with credit balances and no overdrafts (authorizedor unauthorized). Rating methodologiesfor low-defaultloan books are based on expert appraisals. Qualitative and quantitative criteria (comprising the characteristics of the counterparty being rated) are assessed to determinea score and rating for the counterparty.That rating is then linked to the counterparty's Probability of Default (PD), which is calibrated using observed external default data and internal ratings-baseddata. A PD range cannot be quantified due to the low numberof internal defaults. Observed results are backtested in accordance with Group PD and LGD backtesting standards. Statistical indicators are then calculated and qualitative analyses performed on internaldefaults recorded. Dependingon scope, LGD models for low-default books are based on: statistical modeling incorporating internal collections data, where ● the Group has a sufficient data history on observed defaults (e.g. Corporate Unsecured LGD); statisticalmodelingincorporatinginternaldata (e.g. InstitutionLGD ● and Sovereign LGD); stochasticmodelingof collateralvalue distribution(e.g. Specialized ● Financing LGD). In any event, internal defaults are subsequentlybacktested.

Expected Loss Best Estimate. (1)

88

Risk Report Pillar III 2018