This study examined the predictive power of students' demographic characteristics, reading attitudes, school characteristics, and teacher-informed reading activities on three metacognitive reading skills: understanding and remembering, summarizing, and assessing credibility and their influence on 15-year-old students' reading scores. The dataset included 612,004 students in 80 countries who completed the 2018 PISA assessment. We tested three random forest models, each tailored to a specific metacognitive skill, revealing differing fits. Findings indicate that students' self-guided reading strategies and socioeconomic status had highest predictive power, grade repetition and teacher-guided activities showed less impact. These findings have important implications for research, policy, and practice. We apply machine learning to identify best-fitting models for each metacognitive skill and compare the findings with the baseline models.