Abstract We benchmark the performance of widely used long short-term memory (LSTM) models in predicting standardized implied volatility (IV) of equity options against a range of alternative time series models. We forecast option prices over the period 2018-2023 and find universal models that are trained on data pooled across all underlyings to perform the best. The highest forecasting accuracy is achieved by a bidirectional LSTM model with one hidden layer. This provides evidence outside the stock markets that using universal models with pooled data across underlyings to train neural networks significantly improves the forecasting accuracy for equity option prices. Once the machine learning model is trained, one can inexpensively use the trained models to predict option prices up to 7 years into the future.
Fritzsch et al. (Sat,) studied this question.