Artificial intelligence (AI) has gained widespread attention as a potential solution for complex modeling problems. One area where its application has been explored is in managing risks associated with derivative contracts in investment banking. However, despite the positive reports, concerns have arisen regarding its practical applicability. In this article, we critically analyze a recent study published in The Journal of Finance and Data Science that explores whether reinforcement learning (RL) agents can be trained to hedge derivative contracts.
The Challenge of Training Data
The study conducted by researchers from Switzerland and the U.S. acknowledges the significance of training data in effectively training AI agents. The authors highlight the fact that training an AI on simulated market data can yield favorable results only if the simulation accurately reflects real market conditions. However, the data consumption of many AI systems is deemed to be excessive, posing a challenge in terms of collecting sufficient training data.
The Need for Market Simulators
To tackle the shortage of training data, researchers often resort to assuming the existence of an accurate market simulator. However, establishing such a simulator presents a classic financial engineering problem: selecting a model for simulation and calibrating it. This approach essentially aligns AI-based methods with traditional Monte Carlo methods that have been widely used for decades. Furthermore, the study emphasizes that AI can hardly be considered model-free unless there is an abundance of market data available for training, which is rarely the case in realistic derivative markets.
The Role of Deep Contextual Bandits
The collaborative study between IDSIA and UBS, a leading investment bank, is built upon the foundation of Deep Contextual Bandits. These bandits, well-known in RL, are recognized for their data efficiency and robustness. The model developed in the study embraces the operational realities faced by investment firms in the real world, incorporating end-of-day reporting requirements and demanding significantly less training data compared to conventional models. Its adaptability to changing market conditions further sets it apart from traditional approaches.
The senior author of the study, Oleg Szehr, emphasizes the importance of practical considerations in financial institutions. He acknowledges that real-world operational requirements and data availability play a crucial role, overshadowing ideal agent training. The strength of the newly developed model lies in its alignment with the business operations of investment firms, making it a practical solution for risk management.
Although the newly developed model is relatively simple, a rigorous assessment of its performance demonstrates its superiority over benchmark systems in terms of efficiency, adaptability, and accuracy under realistic conditions. The study concludes that, in the realm of risk management, less is often more, highlighting the significance of streamlined and effective approaches.
While AI holds promise in the field of risk management for derivative contracts, it is essential to critically evaluate the practicality and viability of the methods employed. The study analyzed in this article presents a novel approach that addresses some of the concerns related to training data and operational realities. Future research in this area should continue to explore innovative AI-based solutions while considering the practical implications of implementation in real-world financial settings.