New AI method for hedging finance market transactions

Hedging is a form of risk management in which an investment is hedged by a second, opposite transaction. Larger institutional investors in particular, such as banks, insurance companies, pension funds and energy companies, use this method to hedge large transactions on a daily basis in order to control risks and ensure financial stability. ‘More robust hedging strategies can help to limit losses in times of stress and thus indirectly increase the stability of the financial system,’ says Prof. Dr Tobias Sutter from the School of Economics and Political Science at the University of St.Gallen (SEPS-HSG).

A turtle suddenly becomes a rifle

Deep hedging now uses neural networks, a form of artificial intelligence, to optimise such hedging strategies directly from simulated market scenarios, thereby minimising the risk of an investment portfolio. However, there has been a problem with this approach, which first became apparent during the training of AI models for image recognition. ‘When AI learned to recognise animals in images, it became apparent that neural networks can be misled by extremely small changes in individual pixels that are practically invisible to humans,’ explains Prof. Dr. Lukas Gonon from the School of Computer Science at HSG (SCS-HSG). An image that looks unchanged to the human eye is suddenly misclassified by the AI system. ‘A turtle, for example, is suddenly identified as a rifle by the neural network.’ The reason for this is that neural networks base their decisions on very fine statistical patterns in the data. These patterns are stable within the training area, but can abruptly shift outside of it. ‘Although neural networks generalise well, this only applies within the data distribution they have seen during training,’ explains Lukas Gonon.

Unforeseen market data takes AI by surprise

A similar problem was also evident in deep hedging models. These are typically trained either with historical market data or with synthetic data from financial models. ‘However, real markets exhibit jumps, asymmetric movements or sharp trend reversals that are often smoothed out or insufficiently included in model simulations,’ says Lukas Gonon. ‘These differences cause a deep hedging model to learn strategies that are optimal for the simulated or past world, but not necessarily for the current or future market.’

New training method applied

To address this problem, Lukas Gonon, Tobias Sutter and Guangyi He from Imperial College London applied a novel ‘adversarial’ training framework. The model was trained not only with ‘clean’ data, but also with data that was deliberately distorted to simulate the worst-case scenarios of data distribution. ‘The model is deliberately confronted with such “difficult” deviations in order to learn to make robust decisions even outside the idealised training world.’

Better results for greater financial market stability

The results of the extensive numerical experiments are promising. First, the scientists showed that conventional deep hedging models are severely impaired even by minor disturbances in the input distribution. However, the adversarially trained deep hedging strategies consistently outperformed their classical counterparts. In 5,000 samples, adversarial training reduced the average hedging loss by 54% compared to conventional training. Deep hedging models trained in this way could therefore lead to more stable and reliable hedging decisions – a decisive advantage in the often volatile financial markets.