This paper presents an original procedure for selecting the reference online signature instances of a writer, an important issue for any effective signature verifier. To this end, for each signature instance, we propose a novel complexity measure, by exploiting a global description of signatures in the frequency domain as well as a global statistical modelling of each signature instance. To select the reference signatures, we propose a method based on the distribution of complexity values for all the available genuine signatures. The 2500 genuine samples of MCYT-100 online database are used in this study. Experimental results show the effectiveness of the method and of the here proposed complexity measure for this specific task.