Training DNN with PDE and the Heston stochastic volatility model (1993) improves price prediction in terms of the balance between speed and accuracy, compared to standard methods for valuing American options

The adaptation of LSMC using DNN provides an estimation of the American option price with low bias

Accurate training of DNN requires more computational time

RF trained with LSMC always achieves better results than the L-S model

Its simplicity and accuracy make it an ideal algorithm

The Root Mean Square Error (RMSE) of the numerical experiments shows that KNN trained with LSMC is promising for pricing American options

Further research is needed to determine the dimension threshold required for KNN estimators to be viable

Training MLP with LSMC is not much better than the L-S model when there are few exercise dates. However, its efficiency and accuracy in terms of RMSE are better when there are many dates

High efforts and computational resources are required

It is necessary to frequently train MLP in response to financial market conditions

Training CNN with LSMC allows for improving the optimal stopping point performance

The improvement is achieved by transforming historical information into a Markov state, along with extracting features from the CNN layer

The results show that this methodology improves the expected value compared to the L-S model

The RF, KNN, and LGBM models trained with LSMC achieve approximate accuracy in terms of Mean Absolute Error (MAE) with no significant differences

Their combination through stacking increases accuracy in terms of MAE

On average, the price estimates approach their real value

The price estimated by RF trained with LSMC is similar to that obtained through the L-S model, but slightly higher in terms of Mean Squared Error (MSE)

RF generally performs better in the context of nonlinear, highly correlated multidimensional models due to its random tree structure