Input modelling for financial simulations using the bootstrap

Inputs are the fuel that powers simulations. If one is designing a new system or product, inputs are scenarios representing the conditions with which the new system must cope. The need for large numbers of realistic scenarios applies to every domain in which simulation is used as part of the system design process. To be useful, scenarios should mimic the underlying data generating process by reflecting the auto- and cross-correlations of the historical data. We describe a new scenario generation procedure based on the nearest-neighbour bootstrap. We also propose a new performance evaluation criterion for multivariate time-series scenario generators based on the distribution of a composite correlation discrepancy measure. We illustrate the new method and measure by generating simulated scenarios for the US Treasury yield curve.