Reproduce “Encoded Value-at-Risk: A Predictive Machine for Financial Risk Management” on OpenResearch

About this paper

Hamidreza Arian, Mehrdad Moghimi, Ehsan Tabatabaei, Shiva Zamani

Encoded Value-at-Risk (Encoded VaR) introduces a semi-parametric method for financial risk management, leveraging Variational Autoencoders (VAEs) to model complex asset return distributions and dependencies. The approach consistently achieved competitive or superior performance when benchmarked against eleven traditional VaR models across three major stock exchanges, particularly excelling on firm-focused loss functions at 95% and 99% confidence levels.

Key insights

• VAEs can effectively filter noise from financial data; generated returns had 98% of variance explained by signal eigenvectors, a substantial increase from 37% in real data, while preserving 79% of the genuine market signal.
• The model successfully learns and reproduces crucial non-linear dependency structures among assets, evidenced by high eigenvector overlap between real and generated data, without requiring explicit assumptions.
• The generative capability of VAEs provides a computationally efficient Monte Carlo engine to produce thousands of realistic future scenarios that inherently capture stylized facts like fat tails and asymmetry.

Summary by alphaXiv.