Machine learning (ML) has increasingly become a vital tool for accurate pavement condition prediction, enabling more effective maintenance planning and resource allocation. This study applies the random forest (RF) algorithm to predict the pavement condition index (PCI) for 2022 over a short forecast horizon of 4 years across residential and nonresidential (main, collector, and industrial) streets in Skellefteå Municipality, Sweden. The models are based on manually assessed pavement data from 2014 and 2018 and validated against the 2022 pavement condition data. Several performance metrics were applied to evaluate the models’ flexibility and generalization across a broad set of input variables, including pavement age (A), several distress mechanisms (D), previous PCI ratings (S) and average weekday traffic volumes (T). Models for nonresidential streets were further stratified by maintenance surface treatments to assess predictive performance across treatment types, integrating maintenance effects through pavement age and treatment characteristics. Among the residential street models, the RF(A+D+S) achieved the highest accuracy when using the combined dataset (2014 and 2018). Similarly, for nonresidential streets, the RF(A+D+S+T) model performed the best. Furthermore, model performance varied slightly across maintenance treatment categories. The most significant predictors were Status2018 (the PCI rating assessed in 2018) and the weighted distress value (WDV). Use of the combined dataset did not result in a notable improvement in prediction accuracy compared with use of the 2018 dataset alone. Overall, these findings from a geographically constrained dataset suggest that RF models may be useful for PCI prediction and may support data-driven pavement management at the municipal level.