Benchmarking Neural Networks for American Option Pricing

Machine learning techniques have revolutionized the field of financial engineering by providing accurate and efficient methods for pricing American options. This research project aims to explore the effectiveness of deep learning algorithms in accurately pricing American options. The project is divided into two schemes: Scheme I employs a sequence of neural networks, while Scheme II utilizes a single aggregate neural network to eliminate time discretization. By testing various combinations of neural network hyperparameters in both schemes, we seek to optimize the accuracy and computational speed for pricing nine different Put and Call options. Our results are compared against existing efficient algorithms, such as polynomial regression and random forest, as documented in [5]. Based on the analysis of optimal hyperparameters that enhance the accuracy of machine learning-based American option pricing, we identify the top five solvers (hyperparameter sets) in Scheme I and the top two solvers in Scheme II. These solvers are benchmarked and reproducible, serving as reference points for future comparisons with prior studies