Background: Early warning signals (EWS) based on dynamical systems theory, such as increased autocorrelation (AR) and variance, may indicate impending mood episodes in bipolar disorder (BD). This study examines whether smartphone-based digital phenotyping can detect these signals before depressive and (hypo)manic episodes. Methods: We analyzed smartphone sensor and self-report data from 29 BD patients (16 female, mean age of 43,97 +/- 11.9 years) over one year, totaling 10,587 patient days and including 30 depressive and 20 (hypo)manic episodes in 22 patients. 26 bi-weekly expert interviews per patient established daily disease status. AR, variance and moving averages were calculated from passively assessed activity, communication, and smartphone-usage, as well as self-reported sleep parameters. Multilevel logit models assessed whether these measures could predict pre-episode weeks of depression or mania compared to euthymic days. Receiver operating characteristics (ROC) curves evaluated clinical utility. Results: Several parameters significantly predicted pre-episode weeks, but no single robust predictor emerged. Manic episodes were best predicted by altered AR and variance in activity-related measures, while sleep parameters predicted both manic and depressive transitions. Latent factors combining multiple parameters showed stronger predictive potential than individual variables. However, ROC analyses revealed that even the best predictors did not meet predefined clinical utility thresholds. Conclusions: Smartphone-based digital phenotyping holds promise for early detection of BD mood episodes. However, predictive accuracy remains below clinically useful levels. Future research should refine parameters, explore machine-learning approaches, and optimize analytical frameworks to realize the full potential of relapse prediction in BD.