Neural Networks for Delta Hedging Under Market Frictions

Summary

This paper examines whether deep neural networks can learn delta hedging strategies for options when the idealized assumptions behind Black–Scholes—such as continuous trading and no transaction costs—do not hold. It compares recurrent neural networks, temporal convolutional networks, attention networks, and span multilayer perceptrons, and describes NNHedge as a framework for developing and evaluating these models.

The reported experiments fit and assess the models on simulated European call options whose underlying asset follows geometric Brownian motion. The authors report that networks trained to imitate conventional hedging strategies achieved lower profit-and-loss scores, and that the networks relied more on historical data when estimating current delta. These results offer a comparison of architectures in a controlled setting, not evidence of performance in live markets. The document does not establish that learned hedges outperform conventional approaches under real trading conditions, and the simulated price process limits how broadly the findings can be applied.

Key ideas

• The study compares recurrent, temporal convolutional, attention, and span multilayer perceptron networks for delta hedging.
• It tests the approaches on simulated European calls with a geometric Brownian motion underlying.
• The authors report lower profit-and-loss scores when networks are trained to imitate conventional hedging.
• The networks appear to use historical observations more heavily when estimating current delta.
• The simulated setup does not establish how the approaches perform in live markets.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.