Numerical Data-Driven Prediction

Fraunhofer Institute for Algorithms and Scientific Computing SCAI

We develop and investigate algorithms for numerical data-driven prediction. The aim here is to make useful predictions for future behavior of systems based on automatically collected current and historical data. 

Data-driven predictions appear in many application domains like

  • technical trend prediction for foreign exchange rates and other financial products
  • recommendation engines for online shops

The scientific focus is the research and development of efficient numerical
methods for high dimensional problems and is the common ground for the different applications investigated. The developed approaches are based on the modern numerical methods of »sparse grids« and »low rank tensor decompositions« to represent non-linear functions. Their computational effort only scales linearly in the number of data and they achieve an advantageous prediction quality. Therefore they can be powerful even in situations were common non-linear data analysis methods cannot cope with the amount of data. These new tools have the strong potential to be successfully used for the large amounts of data often found in industrial applications.