Not logged in.
Quick Search - Contribution
Contribution Details
| Type | Master's Thesis |
| Scope | Discipline-based scholarship |
| Title | Consistent Scenario Generation of Financial Time Series |
| Other Titles | From Spectral Density Theory to Conditional Adversarial Networks |
| Organization Unit | |
| Authors |
|
| Supervisors |
|
| Language |
|
| Institution | University of Zurich |
| Faculty | Faculty of Business, Economics and Informatics |
| Number of Pages | 87 |
| Date | 2020 |
| Abstract Text | The objective of the Thesis is to investigate novel methods for scenario generation beyond classic Monte Carlo simulation of independent sampled from a multivariate Gaussian distribution. One of the main drawbacks of naive Monte Carlo methods is the difficulty to produce "likely" future scenarios, especially in a high dimensional context. This is due to the poor scalability of Yule-Walker like methods for multivariate Vector Autoregressive Models and the difficulty to insert structural constraint on the auto-cross correlation. Monte Carlo simulations are often performed using very strong assumptions (independent returns, lack of causal structure, etc.), that are likely to be rejected when facing real world data. The topic of scenario generation is of primary importance in quantitative finance, where the ability to obtain more likely and accurate scenarios to compute risk figures and potential profits, maintaining computational feasibility, is undoubtedly a significant desire of many risk and asset managers. Two main approaches are investigated. The first one is based on spectral density theory and can generate multivariate scenarios with a specified auto-cross causal correlation structure. The method applies some modifications to the algorithm proposed in Chambers, 1995, which is currently not publicly implemented in any software library. The second approach is based on a highly customized Deep Learning Architecture, trained in an adversarial setting. This method has the parameter capacity to potentially generate multivariate scenarios with even consistent non linear structure, generalizing the properties of the first algorithm. The analysis is conducted over multivariate time series of financial assets, comprising stock indices, commodity futures and foreign exchange pairs at various frequencies. Algorithms are evaluated in terms of their generative performance, the fidelity of synthetic samples to real world data, and for their computational feasibility in a production environment. |
| Export |
BibTeX
|