This work deals with the derivation of optimal composed intensity measure (Icomp) for the seismic assessment of masonry buildings based on Lasso regression [1]. This methodology is largely used in statistical modelling and machine learning to derive simple regression models with fewer parameters (i.e., shrinking will force some coefficients to become zero). A case study representative of tall and slender residential masonry buildings is modelled in OpenSees [2] using 3D macroelements to account for in-plane (IP) and out-of-plane (OOP) effects [3]. The first three vibration periods are estimated in 0.50 s, 0.42 s, and 0.33 s. Subsequently, probabilistic seismic demand models (PSDMs) are derived as a relationship between intensity measures (IMs) and engineering demand parameters (EDPs) [4]. A total of 100 earthquake motions are selected as input for the non-linear dynamic analyses [5]. Out of 84 measures computed for each accelerogram (i.e., seismological parameters, period and duration-, ground motion dependent-, ground motion dependent compound-, structure independent spectral-, and structure dependent spectral-IMs) the ones with the best correlation within the EPD vs. IM regression are pre-selected to compute the optimal Icomp. The best individual IMs, ranked in descending order, correspond to spectral acceleration, displacement, and velocity at the fundamental period T1, Sa(T1), Sd(T1), and Sv(T1) respectively; Vamvatsikos intensity [6], ("S" _"a" ) ̅; Peak Ground Velocity, PGV; Modified Acceleration Spectrum Intensity [7], ASI*; Cordova intensity [8], Sa*; Velocity Spectrum Intensity, VSI; and Vamvatsikos intensity ("S" _"a" ) ̅ ̅. The proposed Icomp shows large enhancements in terms of conventional statistical indicators of PSDMs [9], i.e., efficiency, practicability, proficiency, and correlation up to R2 = 0.84, for the prediction of the maximum average roof displacement, max(("∆" _"r" ) ̅). Finally, consistent fragility curves obtained from the aggregation of the cloud analysis results are presented to examine the performance of Icomp against individual IMs. It is observed that the collapse prediction is comparable with any of the individual IMs, being the prediction computed through Icomp approximately on the average of all IMs.