Sloppy vs Rigid Parameter Calibration in Complex models
5 May | Model, data, and optimisation approach should be considered as ‘new complex systems. Prof Dr Didier Sornette's talk focusses on making sense of models with multiple parameters and extracting the desired ones.
Model, data, and optimisation approach should be considered as ‘new complex systems’ in need of a general theory. In this talk, Prof Sornette proposes a powerful step in this direction by constructing a natural parametric hierarchy that is ubiquitous for multi-parameter models and which provides a roadmap on how models should be calibrated to experimental data in order to significantly reduce estimation uncertainty.
Employing sophisticated parameter estimation techniques with information from eigenvalues and eigenvectors of the Fisher Information Matrix, the proposed calibration scheme is capable of turning non-identifiable parameters into identifiable ones. Using a GARCH (1,1) model, Monte Carlo simulation results show reduction of estimates uncertainty by a factor of 8, relative to the standard Quasi Maximum Likelihood
approach. The performance of the approach is proven to be good even for small sample sizes (i.e. N=100, …, 200).
Prof Sornette will illustrate the capabilities of this new method to the calibration of sums of exponentials, mixtures of Gaussians, and the
log-periodic power law singularity model of financial bubbles.
Prof Didier Sornette is professor at the Chair of Entrepreneurial Risks at the Swiss Federal Institute of Technology Zurich (ETH Zurich). He is also professor of the Swiss Finance Institute, associated with both the department of Physics and the department of Earth Sciences at ETH Zurich.
Prof Didier Sornette was previously concurrently professor of Geophysics at UCLA, Los Angeles, California and research professor at the French National Centre for Scientific Research, working on the theory and prediction of complex systems.
He is currently principal investigator at the Future Resilient Systems programme, where he leads the research on Resilience Metrics and Outliers