Accurately predicting Active Layer Thickness (ALT) is essential for assessing permafrost stability and mitigating risks to Arctic infrastructure, ecosystems, and communities. Traditional physics-based models struggle with generalization due to sparse in-situ observations and complex environmental drivers. This study introduces a data-driven framework for ALT prediction in the Circum-Arctic combining field measurements, satellite and reanalysis data, and machine learning (ML) techniques. Maximum ALT for a given year is paired with geospatial inputs from earlier years, enabling prediction at four horizons: Same year, +1, +2, and +5 years. A robust multi-stage feature selection strategy is then applied to identify the most informative and temporally stable ALT drivers. To predict ALT, a weighted ensemble of three ML models (Categorical Boosting, Extremely Randomized Trees, and Bootstrap Aggregating) is trained separately for each horizon. The framework is evaluated using data from 115 sites across the Circumpolar Active Layer Monitoring network. Results show that the proposed approach consistently outperforms multiple baseline models across all horizons, with R^2 above 0.8 and RMSE under 25 cm. Future work will be focused on conducting ALT trend analyses to develop a multiclass index of permafrost degradation and regeneration at the Circum-Arctic scale.