Feature-based generators for time series data

A variety of interesting domains, such as financial markets, weather systems, herding phenomena. etc., are characterized by highly complex time series datasets which defy simple description and prediction. The generation of input data for simulators operating in these domains is challenging because process description usually involves high-dimensional joint distributions that are either too complex or simply unavailable. In such applications, a standard approach is to drive simulators with (historical) trace-data, along with facilities for real-time interaction and synchronization. But, limited input data, or conversely, abundant but low-fidelity random data, limits the usefulness and quality of the results. With a view to generating high-fidelity, random input for such applications, we propose a methodology which uses the original data, as a template, to generate candidate datasets, to finally accept only those datasets which resemble the template, based upon parameterized features. We demonstrate the methodology with some early experimental results.