Valuation of American options using machine learning: beyond Longstaff–Schwartz and hybrid models

This work explores the potential of different machine learning (ML) algorithms in the valuation of American options (Aos), contrasting them with the Longstaff–Schwartz (L–S) model. To carry out this research, the algorithms K-Nearest Neighbors (KNN), Random Forest (RF), Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN) are employed using RStudio. The project specifically targets the prediction of the price of Apple’s put Aos through regression, utilizing data extracted from Bloomberg as a case study. To evaluate the model’s performance in a multidimensional context, we use both historical and stochastic volatility. The results show that these ML algorithms achieve a notable improvement in the performance and accuracy of Aos price predictions compared to the L-S model. RMSE values are very similar using historical and stochastic volatility, the most notable difference appearing in the L-S model. Prior trends in the literature show the development of hybrid models, which combine traditional techniques with the predictive capabilities of ML algorithms in the valuation of Aos in a more efficient and accurate way than the L-S model. However, our paper determines whether the supervised training of ML algorithms exclusively with historical financial market data, without relying on traditional methods, achieves better results than those of the L-S model. This ML “stand-alone” approach to price Aos faces the inability to derive hedging strategies and optimal exercise conditions. Future efforts could explore integrating deep reinforcement learning to identify optimal exercise policies or developing hybrid models that combine ML with traditional frameworks.