Historical centers of Europe and Croatia are often formed by unreinforced masonry building aggregates that developed as the layout of the city or village was densified. In these aggregates, adjacent buildings can share structural walls with an older and a newer unit connected either by interlocking or just by a layer of mortar. Observations after for example the recent Central Italy and Croatia earthquakes showed that joints between the buildings were often damaged. This indicated a possible out-of-phase behaviour of units which can lead to the interaction which is demanding to capture with numerical models. The analysis of such building aggregates is difficult due to the lack of guidelines, as the advances were impeded by the scarce experimental data. The SERA project AIMS (Seismic Testing of Adjacent Interacting Masonry Structures) comprised a shake-table test of an aggregate of two buildings under two horizontal components of dynamic excitation, accompanied by the blind prediction competition. Each group was provided with a complete set of construction drawings, material properties, testing sequence and the list of measurements to be reported. After the results were reported, participants were able to compare the results, apply actual accelerations, and update their models within the postdiction phase. The prediction and postdiction of EPFL model were based on an equivalent frame model with a newly developed macroelement able to simulate both the in-plane and out-of-plane behaviour of unreinforced masonry piers, and a newly developed 3D material model allowing to simulate the interaction between the units. This paper deals with the prediction submitted by the EPFL team and discusses the results and possible pitfalls in modelling assumptions leading to unsatisfying prediction. Lessons learned are applied by updating the model for the postdiction analyses and discussing the updated results with the goal to improve the way we model unreinforced masonry aggregates using the equivalent frame approach.