Overcoming the Feature Selection Issue in the Pricing of American Options

The feedforward neural network Monte Carlo method (FNNMC) exhibits more robustness and<br>accuracy than the state-of-the-art least squares Monte Carlo method (LSM) in pricing several<br>American-style options. Specifically, the FNNMC price estimates are accurate for basket options,<br>where the FNNMC price errors are more than four times smaller than the LSM with the best<br>choice of basis functions. By training the neural network the FNNMC avoids the issue of choosing<br>a proper set of basis functions. Hence we circumvent manually engineering the features for each<br>type of option. Furthermore, we explore in-depth the hyperparameter selection for the FNNMC.<br>In the exploration, we use a novel approach called price grid search, where the search is done at<br>the price level instead of at the usual regression level.