Quantum Deep Hedging

Potential of quantum-based machine learning methods as a hedging strategy

The key technology of machine learning / artificial intelligence (AI) has undergone a perceptible revolution in recent years, developing into an important strategic, economic and industrial success factor. The use of AI methods for classification, anomaly identification and regression analysis is increasingly significant in the financial sector, for example in terms of portfolio optimisation strategies. In this context, we will be presenting our analyses and results on utilisation of what is known as “quantum deep hedging”, which describes hedging of financial instruments based on deep neural networks (DNNs) with “reinforcement learning” techniques and elements of quantum neural networks.

Hedging
In general, hedging refers to the strategy of mitigating against, reducing or minimising risks that can occur due to negative price changes in the trading of financial instruments. Common standardised hedging methods include delta or gamma hedging, as well as Monte Carlo simulation based methods. Developing a successful hedging strategy is extremely challenging. As well as the natural dynamics of the market and the numerous correlations between different instruments, influences from secondary conditions such as liquidity, limitation of capital and transaction costs add to the complexity of the problem.

Reinforcement Learning
Machine learning systems can essentially be classified based on particular criteria, such as the type of supervision. Alongside supervised or unsupervised learning, what is known as “reinforcement learning” represents a separate methodology, in which the learning system (also known as the “agent”) interacts with its environment and takes actions based on certain sets of rules, for which it receives either a positive or negative reward (or penalty). The rules are updated according to the learning steps until an optimum strategy has been identified. The aim of the agent is to learn a behaviour that maximises its expected rewards.

Deep Hedging
Deep hedging involves modelling trading decisions within the hedging strategy using a deep neural network, enabling different additional factors such as past trading decisions, liquidity restrictions or news analysis to be integrated into the overall model as “features”. During calibration of a stochastic pricing model that acts as an input for the neural network, either the implicit volatility method or a classic time series method based on historic value development can be used. The aim of the neural network learning process is to minimise a defined mathematical objective function within the formulation of a mathematical optimisation problem. In the case examined here, we use the 10% expected shortfall as the risk measure. […]

To continue reading, please download the full article above.

The article was published in the 2024 yearbook of The Frankfurt Institute for Risk Management and Regulation (FIRM).