The log-linear probability model has been initially introduced by Nerlove and Press (1973) for the analysis of contingency tables constructed from business survey data. We extend this modelling approach to the dynamic analysis of multivariate qualitative processes with the application to technical analysis of financial returns in mind. We develop the dynamic qualitative models with pairwise and/or three-wise interactions, discuss the interpretations of the interaction parameters,study the filtering and prediction algorithms, and compare the approach to machine learning models as the restricted Boltzmann machine and the normalizing flows.