Generative Adversarial Networks applied to synthetic financial scenarios generation

• Published in: Physica A Vol. 623; p. 128899
• Main Authors: Rizzato, Matteo, Wallart, Julien, Geissler, Christophe, Morizet, Nicolas, Boumlaik, Noureddine
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2023; Elsevier
• Subjects: Applications; Computational Finance; Conditional data augmentation; Deep neural networks; Financial scenarios; Generative Adversarial Networks; Machine Learning; Quantitative Finance; Risk management; Statistics; Time series generation; Generative Adversarial Networks; Deep neural networks; Conditional data augmentation; Risk management; Financial scenarios; Time series generation; Data Augmentation; Deep neural networks Generative Adversarial Networks Conditional data augmentation Financial scenarios Risk management Time series generation
• ISSN: 0378-4371; 1873-2119; 0378-4371
• DOI: 10.1016/j.physa.2023.128899

In this paper, we introduce Jinkou, a GAN-based algorithm that allows for the conditional generation of synthetic multivariate time series. The set of variables whose distribution is to be replicated include specific variables taking different values for different objects, as well state variables describing the state of the world, common to all objects at a given date and potentially influential on the specific features. The conditioning process is specified at inference time, and only involves state variables; it simply consists in setting lower and/or upper bounds on their values. The generative model is trained as an un-conditioned generator and is agnostic of any scenario the user might set at inference time. The use case considered in this pilot study is of interest for the financial industry: the generator produces random samples of the instrument-specific features over time (e.g their price, size or the risk for securities). Such generation is conditioned on user-defined macroeconomic assumptions/scenarios involving global variables, such as inflation, oil prices or interest rates. We introduce numerical metrics to assess the statistical closeness between the two multivariate distributions of historical and artificial data. As proof of concept, we test the proposed algorithm by reproducing the value variation for two possible portfolios, Energy and Financial, conditioned on scenarios for which a consensus is present in the community. Jinkou allows us to recover some classical stylized facts about the financial markets, this ability constituting a proof of its efficiency. [Display omitted] •We synthesize data conditioned over continuous values not known at training time.•Jinkou allows us to recover the most known stylized facts in financial markets.•No algorithmic details are specific to the use case considered in this study.