Key words: Multivariate time series, nearest neighbor bootstrap, scenario generation
Scenario Generation for Multivariate Time Series Data using the k-Nearest-Neighbor Bootstrap
Haiyan Huang
Thomas R. Willemain
Department of Decision Sciences and Engineering Systems
Rensselaer Polytechnic Institute
Due to the scarcity of input data and the limitations of existing scenario generation approaches, simulation modelers face a dilemma: either use a very small number of realistic input scenarios or use a large number of unrealistic scenarios. This dilemma would disappear if we were able to generate input scenarios that were simultaneously numerous and realistic. Bootstrapping is a natural approach for this problem. However, the autocorrelation and crosscorrelation structures in multivariate time series data present challenges to conventional bootstrap methods. The objective of this research is to generate artificial multivariate time series data as input scenarios for simulation models using bootstrapping.
I will introduce the k Nearest Neighbor Bootstrap (NNB) for generating multivariate time series data. Applications in financial stress testing and risk assessment will be addressed.