Adaptive value-at-risk policy optimization: a deep reinforcement learning approach for minimizing the capital charge

• Main Author: Banhudo, Guilherme Sousa Falcão Duarte
• Publication Date: 2019
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10071/19197
• topic: Value at risk; Basel accord; Artificial intelligence; Deep learning; Deep reinforcement learning; Proximal policy optimization

In 1995, the Basel Committee on Banking Supervision emitted an amendment to the first Basel Accord, allowing financial institutions to develop internal risk models, based on the value-at-risk (VaR), as opposed to using the regulator’s predefined model. From that point onwards, the scientific community has focused its efforts on improving the accuracy of the VaR models to reduce the capital requirements stipulated by the regulatory framework. In contrast, some authors proposed that the key towards disclosure optimization would not lie in improving the existing models, but in manipulating the estimated value. The most recent progress in this field employed dynamic programming (DP), based on Markov decision processes (MDPs), to create a daily report policy. However, the use of dynamic programming carries heavy costs for the solution; not only does the algorithm require an explicit transition probability matrix, the high computational storage requirements and inability to operate in continuous MDPs demand simplifying the problem. The purpose of this work is to introduce deep reinforcement learning as an alternative to solving problems characterized by a complex or continuous MDP. To this end, the author benchmarks the DP generated policy with one generated via proximal policy optimization. In conclusion, and despite the small number of employed learning iterations, the algorithm showcased a strong convergence with the optimal policy, allowing for the methodology to be used on the unrestricted problem, without incurring in simplifications such as action and state discretization.