Major Depressive Disorder (MDD) is highly prevalent and characterized by often debilitating behavioral and cognitive symptoms. MDD is poorly understood, likely due to considerable heterogeneity and self-report-driven symptomatology. While researchers have been exploring the ability of machine learning to screen for MDD, much less attention has been paid to individual symptoms. We posit that understanding the relationship between objective data streams and individual depression symptoms is important for understanding the considerable heterogeneity in MDD. Thus, we conduct a comprehensive comparative study to explore the ability of machine learning to predict nine self-reported depressive symptoms with call and text logs. We created time series from the logs of over 300 participants by aggregating communication attributes- average length, count, or contacts- every 4, 6, 12, or 24 hours. We were most successful predicting movement irregularities with a balanced accuracy of 0.70. Further, we predicted suicidal ideation with a balanced accuracy of 0.67. Outgoing texts proved to be the most useful log type. This study provides valuable insights for future mobile health research aimed at personalizing assessment and intervention for MDD.