Monte Carlo Simulation with Machine Learning for Pricing American Options and Convertible Bonds

Li, Szepesvari and Schuurmans (2009) show that reinforcement learning (RL) algorithms are superior to the classical methods (such as Longstaff and Schwartz (2001)) in pricing American options using Monte Carlo simulation. We extend their techniques to the problem of pricing convertible bonds and show that RL outperforms LS on this task. Additionally, we propose a new method, based on the random forest algorithm from machine learning [Breiman (2001)], that can be used for pricing both American options and convertible bonds with Monte Carlo simulation. We show that this algorithm outperforms LS and is also superior to RL in most cases. We demonstrate how to use Monte Carlo simulation with the methods described above for pricing a complex convertible bond trading at the Tel Aviv stock exchange. Like many Israeli convertibles, this bond exhibits the "gradually diminishing principal" feature, meaning that instead of one payment of the principal at maturity, there are multiple principal payments during the lifetime of the bond. This feature presents a challenge to existing models. We also model other exotic features of this bond, such as path-dependent conversion ratio and exchange rate indexation. The prices that we obtain using this model are close to the market prices of the bond.