I got occasions clustered in students, each student has its own number of occasion because naturally ended the test at different time. I don´t know if multievel analysis will be the best, and neither growth curve modeling because i´m interested in the prediction of perfomance by anxiety trajectories, not trajectories itself.

latent growth curve modelling will pull out groups of students who show similar trajectories of anxiety, which you can then characterize using regression. You need MPlus or similar software that does longitudinal SEM-based analyses.

I second what Richard mentioned. Latent growth curve modeling is a great option for identifying groups with similar anxiety trajectories. It can be used as an exploratory technique if you do not have specific hypotheses about what kinds of trajectories will emerge and how these trajectories are likely to predict performance.