Providing accurate estimates of time-to-resolution of legal cases contributes to building trust in judicial systems. However, this task is challenging due to the temporal variability of legal proceedings and the multitude of factors influencing the time to resolution. While machine learning methods for remaining-time prediction have been extensively studied in the context of business processes, their application to judicial processes remains underexplored. A key component of remaining time prediction methods is the set of features used to train the machine learning models. In this study, we investigate the predictive power of different feature sets for estimating the time-to-resolution of legal cases in the specific context of Italian courts. We conduct an ablation study to evaluate the contribution of five categories of features: control-flow (activity sequences), temporal context, case attributes, global process state, and judge workload. The results show that, while case attributes alone perform poorly, combining them with state and temporal context significantly enhances the predictive performance. This combination achieves a mean absolute error (MAE) within an acceptable range for practical use, underscoring the value of integrating procedural and contextual information in predictive models for judicial processes.