2018 REGISTRATION DOCUMENT AND ANNUAL FINANCIAL REPORT

2018 Registration document and annual fi nancial report - BNP PARIBAS350

5 RISKS AND CAPITAL ADEQUACY PILLAR 3

Credit risk

Modelled parameter Portfolio

Number of models Model and methodology

Number of years default/loss data Main asset class

CCF/EAD CCF for corporates, banks and sovereigns 1

Quantitative calibrated on internal data > 10 years

Central governments and central banks

French Retail Banking Retail 4 Quantitative calibrated on internal data 10 years Retail

Personal Finance France 1 Quantitative calibrated on internal data > 10 years Retail

Belgian Retail Banking Professionnals & SME 1

Quantitative calibrated on internal data > 10 years Retail

Belgian Retail Banking Individuals 2

Quantitative calibrated on internal data > 10 years Retail

Banca Nazionale del Lavoro Retail Individuals 1 CCF = 100% - Retail

BACKTESTING Each of the three credit risk parameters (PD, LGD, CCF/EAD) is backtested and probability of default benchmarked annually to check the system s performance for each of the Bank s business segments. Backtesting consists in comparing estimated and actual outcomes for each parameter. Benchmarking consists in comparing the parameters estimated internally with those of external organisations.

For the non-Retail IRBA scope, all ratings, including default ratings 11 and 12, for all counterparties to which the Bank has a credit risk exposure, have been recorded over a long period of time. Likewise, observed losses on defaulted exposures during this period are also archived. Backtesting is performed on the basis of this information for each of the risk inputs, both globally and across the scope of each rating model. These exercises aim to measure overall performance and the performance of each rating method, and in particular, to verify the model s discriminatory power (i.e. the less well rated counterparties ought to default more often than the better rated ones) as well as the predictive, conservative nature of the parameters. For this purpose, observed losses and default rates are compared with estimated Global Recovery Rates and Probability of default for each rating. The through the cycle or downturn nature of these ratings and loss rates in the event of default (LGD) is also verifi ed.

For benchmarking work on non-Retail exposures, internal ratings are compared with the external ratings of several agencies based on the mapping between internal and external rating scales. Around 10% of the Group s corporate clients have an external rating and the benchmarking studies reveal a conservative approach to internal ratings.

Performance measurements are also carried out on sub-scopes of homogeneous asset classes for Retail portfolios. If the predictive power or the conservative nature of a model has deteriorated, the model is recalibrated or redeveloped as appropriate. These changes are submitted to the regulator for approval in line with the regulations. Pending implementation of the new model, the bank takes conservative measures to enhance the conservatism of the existing model where necessary.

Backtesting of Loss Given Default is based mainly on analysing recovery cash fl ows for exposures in default. When the recovery process is closed for a given exposure, all recovered amounts are discounted back to the default date and then compared to the exposure amount. When the recovery process is not closed, the future recoveries are estimated by using either the amount of provisions, or historically calibrated statistical profi les. The recovery rate determined in this way is then compared with the initially forecasted rate one year before the default occurred. As with ratings, recovery rates are analysed on an overall basis and by rating policy and geographical area. Variances are analysed taking into account the marked bimodal distribution of recovery rates.

All of this work is reviewed annually in the Capital Committee (see section 5.2 under Capital management). Backtesting is also certifi ed internally by an independent team and the results sent to the Supervisor.

The following two tables present an overview of the performance of models for regulatory risk parameters (PD and LGD) within the context of the Group s IRBA scope, using the following indicators:

■ arithmetical average of the PD: average probability of default of performing loans weighted by the number of number of obligors in the portfolio in question;

■ historic average default rates: average annual default rate (number of obligors defaulting during a fi nancial year relative to the number of performing obligors at the end of the previous year) observed over a long historical period (see Table 30: Main models: PD, LGD and CCF/EAD);

■ arithmetical average of the estimated LGD: average rate of loss in the event of default weighted by the number of obligors or by the amount of EAD depending on the portfolio in question;

■ arithmetical average of the historic LGD observed: the rate of loss in the event of default observed over a long historical period (see Table 30: Main models: PD, LGD and CCF/EAD).