Scenario generation by selection from historical data

Abstract

In this paper, we present and compare several methods for generating scenarios for stochastic-programming models by direct selection from historical data. The methods range from standard sampling and k-means, through iterative sampling-based selection methods, to a new moment-based optimization approach. We compare the models on a simple portfolio-optimization model and show how to use them in a situation when we are selecting whole sequences from the data, instead of single data points.