Flexible Models for Large Financial Micro Datasets

Purpose and goal

The main aim of the project was to exploit the large micro datasets that are nowadays available for analysis to estimate more complex and feature rich models. The project has proposed several new flexible models that allows for nonlinearity and more realistic model assumptions, such as the dynamic mixture model. Another major contribution are new MCMC algorithms for so called big data problems based on efficient sampling of data subsets. These algorithms make it possible to estimate complex models on very large datasets (so called big data) within reasonable time frames.

Results and expected effects

The new models developed in the project has shown to provide useful insights and to produce more accurate predictions. A large portion of the work has been devoted to ensure the practical estimation of complex models on very large datasets, with the ambition of making the models accessible to practitioners. The developed MCMC methods are breaking new ground and will most certainly generate new research within this currently very active research field.

Approach and implementation

The development of models and algorithms was motivated by real data on bankruptcy of Swedish corporations. Previous research has shown clear evidence of nonlinearities, and motivated us to propose new models with interesting interpretation and improved forecasting performance. The complexity of the models made it clear that new estimation algorithms were needed to perform the necessary computations in realistic time. Much of the project was then devoted to the currently very active field of research on algorithms for analysing so called big data.