Deep learning of value at risk through generative neural network models: The case of the Variational auto encoder

Area of Science:

• Computational finance
• Machine learning applications in finance
• Risk management

Background:

• Traditional Value at Risk (VaR) estimation often relies on distributional assumptions.
• Accurate VaR estimation is crucial for financial risk management.
• Generative models offer potential for flexible financial modeling.

Purpose of the Study:

• To introduce a novel method for computing Value at Risk (VaR) using generative neural networks.
• To assess the performance of a Variational Auto Encoder with an energy kernel for VaR estimation.
• To demonstrate a method that does not require prior assumptions on return distributions.

Main Methods:

• Utilized generative neural networks, specifically a Variational Auto Encoder (VAE).
• Employed an energy (Radon-Sobolev) kernel within the VAE architecture.
• Applied the method to estimate Value at Risk for financial assets.

Main Results:

• The proposed method provides an effective estimator for Value at Risk.
• Results align with intuitive expectations and established classical methods.
• Demonstrated good practical behavior in financial asset analysis.

Conclusions:

• Generative neural networks offer a viable alternative for VaR estimation.
• The VAE with an energy kernel successfully estimates VaR without distributional constraints.
• The method shows practical utility and aligns with financial intuition.