| Investment Knowledge |
 |
 |
 |
 |
|
Counterparty Credit Risk Backtesting The methodologies we discuss are based on comparison between the internal model forecasted probability distribution of exposure at various time horizons (calculated for representative counterparty portfolios) and the actual exposures that would have occurred on each portfolio at each time horizon by using in computation historical data on movements in market risk factors. A forecast distribution of market risk factors or exposures has a number of properties. Forecasts are initialised at a particular point in time. The initialisation point is the date and time that a forecast is launched or issued. Each forecast distribution has a time horizon, the time between initialisation and the realisation of the forecast. A forecast initialised on January 1st that realises on January 15th has a 14 day time horizon, a two week forecast. Note that forecasts with different time horizons can have the same initialisation date, ie two week and four week forecasts that realise on 15th and 29th January respectively would both have been initialised on the same date, 1st January. Backtesting is a statistical test with the significance of any result depending on the amount of data used. A backtesting data set is a set of forecasts and the corresponding realisations of those forecasts, ie what actually occurred. This backtesting data set can be put together in a number of ways. Backtesting using data from a single counterparty over a short period of time may not produce a meaningful conclusion about the quality of the EPE models and its sub-components used to generate that exposure. Firms with advanced model permission have addressed the data requirement problem by aggregating backtesting data across a number of dimensions. The possible dimensions are discussed below. The backtesting data set can be aggregated over time, over trades/risk factors or over both time and trades/risk factors. The time period over which data is aggregated is referred to as the observation window. There are a number of methodologies for generating a backtesting data set over a given observation window. A selection of frequently used methodologies are set out below.
|