This study investigates the use of formulaic sequences (FS) in academic writings of Chinese learners of English as a foreign language across different levels of studies at a public university in China. Frequency-based formulas were retrieved from a corpus of student academic texts written at five time points between Year 1 and Year 3. The structure, semantic transparency (as measured by mutual information, abbreviated as MI), function of the formulas and first language -- second language (L1-L2) congruency were explored to identify factors influencing FS development over time using generalized linear mixed-effects modeling (GLMM). Results indicate that MI, structure and assumed learners' proficiency (time points) and their interactions produced significant effects on the development of L2 FS, but function and congruency did not. The individual variation and nonlinearity of FS development were interpreted with Complex Dynamic Systems Theory (CDST). Based on the longitudinal study, we highlight the usefulness of GLMM in accounting for both systematicity and individual variation in L2 development from a CDST perspective.