Languages encode deictic information in their demonstrative systems, but exactly which information is encoded is a matter of variation. The present dissertation explores this variation, with special attention to Romance demonstrative systems, and does so from both a synchronic and a diachronic perspective. Synchronically, the focus is on how the attested cross-linguistic differences can be formalised in featural and, more broadly, syntactic terms. This line of investigation results in the proposal of a novel internal structure for demonstrative elements that ties together a lower person-based component and a higher spatial-based one, overcoming the classic dichotomy between the two and affording the system a larger empirical coverage. Diachronically, the focus is on how the inventory of contrastive demonstrative forms changes in the different (micro-)diachronic stages of a given language, and more concretely on how it shrinks and why. Interestingly, this change only involves demonstrative forms, but not other deictic categories (personal pronouns, possessive forms, etc.). Based on novel generalisations concerning the patterns of change attested across Romance demonstratives and on the conclusions drawn from the synchronic investigations, this work proposes that larger demonstrative systems are unstable because of their featural complexity. The latter hinges on a bias towards monotonic derivations, which triggers feature loss and results in a smaller inventory of demonstrative forms. Additionally, a structural condition on feature loss is identified (as formalised in the Last in–First out principle), which accounts for the concrete patterns of reduction and for the asymmetry between demonstrative systems (unstable) and other indexical systems (stable).