Fueling Our Models: Generation of Probabilistic Scenarios
Wednesday, April 14, 10-10:40am EDT
OR models supporting business decisions often require as inputs massive numbers of probabilistic scenarios depicting future operating conditions. For example, with business operating at ever-lower levels of aggregation and ever-higher frequencies, demand planning and inventory optimization now use models fueled by scenarios representing the randomness of product demand at a daily scale. Even non-decision tasks such as operator training benefit from large numbers of realistic training scenarios. We will review how such scenarios are used, evaluated, and automatically generated using time-series bootstrapping. These results derive from three sources: a framework for evaluating scenario generators (quantity, cost, fidelity and variety), a novel quantitative metric of the quality of bootstrap-based scenarios, and on empirical studies of the ability of trained observers to distinguish between real time series and bootstrap replicates.