systems to consume and deliver data is to leverage in-memory. In-memory views of data canin principle also include transformations. The benefit is the potential speed of data deliveryand that a physical duplicate of data is prevented.Another factor that affects the choice of distributed calculation and storage on a grid ofmultiple computer nodes is the network bandwidth. Network bandwidth is a scarce resourceeven when a dedicated network is set up for the grid. Nowadays multigigabytes of data canbe transmitted among computers. However, oftentimes multiple nodes share one networkcable. If a lot of data are transmitted through the network, it can easily saturate the net-work bandwidth capacity. Therefore, one should consider the network factor in risk systemarchitecture.MODEL RISK MANAGEMENTRisk analytics is the core of a risk system. It is also this book’s focus. In a book that focusesheavily on risk analysis models it is appropriate to discuss model risk management even beforegoing into detailed discussion of any models.A risk system contains many different models used in the risk calculations. Models havetheir lifecycles and need to be frequently validated and tested. The need for more risk calcula-tions and their growing complexity also means that banks need to manage and validate moreand more models that are deployed in the risk system(s). Sometimes models are managed ona firmwide level even if certain risk calculations are executed on a silo risk system level. Afirmwide model risk management approach is motivated by the firmwide stress test require-ments from CCAR and EBA. Given the many different models that are involved in banks’regulatory calculations, banks’ practice in model risk management is also a serious concernfor regulators. For example, in 2011, the US Federal Reserve and the Office of the Comptrollerof the Currency (OCC) issued joint guidance on model risk management practices.
18INTRODUCTIONModel risk can arise from many points in an analytical risk process. It cannot be com-pletely avoided but it can be managed. Model risk can arise and thus ought to be managedin the following aspects:■Model data:The first source of model risk is data. We are familiar with the garbage-in-garbage-outrule. When the input data have major flaws, even the best model can generate mis-leading results. Data flaws can occur not only due to the data quality but also from apartial selection of the data history. For example, in the years preceding the 2007 finan-cial crisis, housing price statistics displayed upward trending for many years with onlysporadic default losses. When banks used this part of the historical data in the hous-ing price models and the credit default risk models it was impossible to obtain tail risklevels anywhere close to the losses in the 2007 financial crisis. Even if data are avail-able that cover a full business cycle with both up- and downturns it is still justoneobservation of a business cycle. The future business cycle swings will most likely not besimilar to the last one. This observation is relevant not only in the context of using datafor model calibration such as how probability of default depends on the business cyclebut also in selection and design of (macroeconomic) stress scenarios for the portfolios.