There is increasing evidence that fine-grained aspects of student performance and interaction within educational software are predictive of long-term learning. Machine learning models have been used to provide assessments of affect, behavior, and cognition based on analyses of system log data, estimating the probability of a student's particular affective state, behavior, and knowledge (cognition). These measures have (in aggregate) successfully predicted outcomes such as performance on standardized exams. In this paper, we employ a different approach of relating interaction patterns to learning outcomes, using dynamical methods that assess patterns of fine-grained measures of affect, behavior, and knowledge as they occur across time. We use Hurst exponents and Entropy scores computed from assessments of affect, behavior, performance, and knowledge acquired from 1,376 middle school students who used a math tutoring system (ASSISTments), and analyze the relations of these dynamical measures to the students' end-of-year state test (MCAS) performance. Our results show that fine-grained changes in affect, behavior, and knowledge are significantly related to and predictive of their eventual MCAS performance, providing a new lens on the dynamic and nuanced nature of student interaction within online learning platforms and how it affects achievement. [For complete proceedings, see ED560503.]