Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Value at Risk Estimation with Generative Adversarial Networks

Tobjörk, David LU (2021) STAN40 20201
Department of Statistics
Abstract
Risk is of large importance for financial institutions and there are many different measures that can be used. A popular one is value at risk (VaR), which is the maximum likely loss for a portfolio of financial assets. Different methods of estimating it has been suggested, one often described is the variance-covariance method. Here a method based on a generative model called generative adversarial networks (GANs) was considered. More precisely, a type of model presented as a way of improving training of GANs called WGAN-GP was used. By simulating synthetic returns for the portfolio VaR was estimated. Two types of models using WGAN-GP was considered, a model that can generate returns conditioned on recent returns and unconditional models.... (More)
Risk is of large importance for financial institutions and there are many different measures that can be used. A popular one is value at risk (VaR), which is the maximum likely loss for a portfolio of financial assets. Different methods of estimating it has been suggested, one often described is the variance-covariance method. Here a method based on a generative model called generative adversarial networks (GANs) was considered. More precisely, a type of model presented as a way of improving training of GANs called WGAN-GP was used. By simulating synthetic returns for the portfolio VaR was estimated. Two types of models using WGAN-GP was considered, a model that can generate returns conditioned on recent returns and unconditional models. The models were backtested and compared to a benchmark model using the variance-covariance approach. When evaluating the models it's shown that the conditional model is very sensitive to changes in the conditioning data. The unconditional models however, show similar estimates to the benchmark model over time. But, only the benchmark model is not rejected for the overall backtesting period by Kupiec's POF test. (Less)
Please use this url to cite or link to this publication:
author
Tobjörk, David LU
supervisor
organization
course
STAN40 20201
year
type
H1 - Master's Degree (One Year)
subject
keywords
generative adversarial networks value at risk finance machine learning neural networks
language
English
id
9042741
date added to LUP
2021-06-29 07:57:54
date last changed
2021-06-29 07:57:54
@misc{9042741,
  abstract     = {{Risk is of large importance for financial institutions and there are many different measures that can be used. A popular one is value at risk (VaR), which is the maximum likely loss for a portfolio of financial assets. Different methods of estimating it has been suggested, one often described is the variance-covariance method. Here a method based on a generative model called generative adversarial networks (GANs) was considered. More precisely, a type of model presented as a way of improving training of GANs called WGAN-GP was used. By simulating synthetic returns for the portfolio VaR was estimated. Two types of models using WGAN-GP was considered, a model that can generate returns conditioned on recent returns and unconditional models. The models were backtested and compared to a benchmark model using the variance-covariance approach. When evaluating the models it's shown that the conditional model is very sensitive to changes in the conditioning data. The unconditional models however, show similar estimates to the benchmark model over time. But, only the benchmark model is not rejected for the overall backtesting period by Kupiec's POF test.}},
  author       = {{Tobjörk, David}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Value at Risk Estimation with Generative Adversarial Networks}},
  year         = {{2021}},
}