Mathematical and Machine Learning Approaches to Value at Risk

Summary

The document collects suggestions for studying the mathematical foundations and newer approaches to Value at Risk (VaR). It mentions combining multiple VaR methods with artificial neural networks, with the network balancing weaknesses among component approaches. It also points to work on correcting the Cornish–Fisher expansion using response surface methodology, with applications to VaR and Conditional Value at Risk, and recommends a practical overview of methods used for VaR calculations.

These are reading leads rather than a tutorial or comparison of model performance. The discussion offers no equations, implementation details, empirical results, or assessment of when the cited approaches outperform simpler alternatives. One response cautions that machine learning may be excessive for a relatively simple and contested risk measure, and another warns readers to assess fashionable methodological labels critically. The material is useful for identifying research directions, but readers would need to consult the cited papers and talk to evaluate the methods and their assumptions.

Key ideas

• Neural networks have been proposed to combine VaR estimates from multiple methods.
• A cited study applies a corrected Cornish–Fisher expansion and response surface methodology to VaR and CVaR.
• The document recommends consulting a practical overview of VaR calculation methods.
• The suggestions are references rather than evidence of comparative performance, and advanced methods warrant critical scrutiny.

Tags

# Value-at-Risk theory papers


# Value-at-Risk theory papers












I am looking for some papers related to the value-at-risk theory. I would like to focus on mathematical aspects of VaR. I would like to read something about modern approaches to VaR (maybe using Machine Learning methods of something like that).

## Answer by Fokko (score 1)

https://quant.stackexchange.com/a/44623

My first thought was that machine learning is quite exaggerated for a rather simple (but often controversial) measure like VaR.

However, after some research I found an interesting paper re the improvement of VaR by combining different approaches under the use of artificial neural networks to balance the downsides of the methods out. So maybe you find this interesting as well: Value-at-Risk Model Combination Using Artificial Neural Networks by Yan Liu

## Answer by Dimitri Vulis (score 1)

https://quant.stackexchange.com/a/44625

I have read the article "Computation of the corrected Cornish–Fisher expansion using the response surface methodology: application to VaR and CVaR" by Charles-Olivier Amédée-Manesme, Fabrice Barthélémy, Didier Maillard (Annals of Operations Research, 2018, https://doi.org/10.1007/s10479-018-2792-4 ). You definitely should read this in order to keep up with the current state of progress, but don't turn off your critical thinking skills when you see keywords like "response surface" and "machine learning".

Also this talk https://cornell.mediasite.com/Mediasite/Play/0db1659795f6472cbc8fb6030f7eac041d by Jonathan Schachter is a very good summary of what people actually use in practice for VaR calculations.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.