Mood episodes in bipolar disorders (BD) are typically associated with changes in sleep and activity patterns. In this study, we present a novel correlation-based approach to examine the structure of relationships between actigraphy-derived variables and clinical status in individuals with BD. Using a large-scale longitudinal study spanning over 2 years of actigraphic recordings from 115 patients with bipolar disorder (BD), we compared three aggregation approaches: mean values, temporal variability, and inter-feature correlation structure, to classify mood states in a strict validation scenario and predict future episodes. Two binary classification subsets (mania–remission and depression–remission) were constructed based on automatically classified labels using clinician-rated scales (Montgomery-Åsberg Depression Rating Scale and Young Mania Rating Scale) and weekly self-assessments. Support Vector Machine models with Radial Basis Function kernels were trained using 7-day windows and forward feature selection in a nested 5-fold cross-validation setup. The classification model achieved statistically significant balanced accuracy: 58% (p < 0.05) for mania–remission and 59% (p < 0.001) for depression–remission. For mania, the most predictive aggregates were shorter mean sleep duration and deviations in 10-hour peak activity across different weeks, while for depression, increased inter-daily variability and intra-daily activity fluctuations emerged as key indicators. The models differentiated future episodes from remission solely from actigraphy data with above-chance accuracy, revealing distinct behavioral signatures for mania and depression. While structural changes in correlation patterns between features differed across mood states, they did not outperform mean or variability-based metrics in classification performance. However, modest classification performance and high inter-individual variability suggest that personalized modeling approaches may be essential for clinically meaningful prediction. Notably, inter- and intra-weekly feature changes provided the strongest predictive signals, suggesting that straightforward alterations in rest-activity rhythms may better reflect clinical episodes than more complex structural metrics.