Predicting opponents' moves and hidden states is important in imperfect information games. This paper describes a method for building a Mahjong program that models opponent players and performs Monte Carlo simulation with the models. We decompose an opponent's play into three elements, namely, waiting, winning tiles, and winning scores, and train prediction models for those elements using game records of expert human players. Opponents' moves in the Monte Carlo simulations are determined based on the probability distributions of the opponent models. We have evaluated the playing strength of the resulting program on a popular online Mahjong site “Tenhou”. The program has achieved a rating of 1718, which is significantly higher than that of the average human player.