Improving implied volatility forecasts for American options using neural networks

This paper explores the application of neural networks to improve pricing of American options. Focusing on both American and European options on the S&P 100 index from January 2016 to August 2023, we integrate neural networks to model the difference between market-implied and model-implied volatilities derived from the Black-Scholes and Heston models. We also employ a pure neural network which does not use pricing models. Our study compares two neural network training architectures: independent training, where the network is trained anew daily, and sequential training, which uses the previous day’s network setup as starting point to retrain the network. The findings reveal that a pure neural network with sequential training significantly reduces the root mean square error (RMSE) of implied volatility predictions compared to traditional pricing models. The inclusion of European option data in the training can enhance the models’ forecasting performance, though its effectiveness varies depending on market conditions.