In vehicles, dozens of electronic control units are connected to one or more controller area network (CAN) buses to exchange information and send commands related to the physical system of the vehicles. Furthermore, modern vehicles are connected to the Internet via telematics control units (TCUs). This leads to an attack vector in which attackers can control vehicles remotely once they gain access to in-vehicle networks (IVNs) and can discover the formats of important messages. Although the format information is kept secret by car manufacturers, CAN is vulnerable, since payloads are transmitted in plain text. In contrast, the secrecy of message formats inhibits IVN security research by third-party researchers. It also hinders effective security tests for in-vehicle networks as performed by evaluation authorities. To mitigate this problem, a method of reverse-engineering CAN payload formats is proposed. The method utilizes classification algorithms to predict signal boundaries from CAN payloads. Several features were uniquely chosen and devised to quantify the type-specific characteristics of signals. The method is evaluated on real-world and synthetic CAN traces, and the results show that our method can predict at least 10% more signal boundaries than the existing methods.