In a data-availability scenario in which a sufficiently long series of qualitatively and quantitatively adequate runoff data covering the recent past is not available for assessing a project’s viability from water-availability assessment, mathematical models can be employed to synthetically generate such a series by using the data of a distant past, if available. However, the development of different categories of mathematical models of varying complexities, particularly with the advent of Machine-Learning (ML), has rendered the task of choosing a suitable model for a data-scarce project challenging. In this study, five system-theoretic Linear-Relation-Based (LRB) and four data-driven ML-Based (MLB) models were applied on past records of ten-daily average runoff and a concurrent derived series of ten-daily rainfall with the objective of identifying a suitable model that could be used for updating the water-availability of the northeast India’s Kamala HEP of 1800 MW that lacks runoff data of more than a decade of the recent past. It emerged from this study that a parsimonious LRB model can perform better than its relatively complex MLB counterparts, and that Occam’s Razor holds in the case of rainfall-runoff modelling, particularly with data of a relatively coarse timestep. These observations substantiate the utility of parsimonious and simpler models in keeping at bay the problems of model overfitting, equifinality and suboptimal performance of relatively complex and non-parsimonious MLB models for rainfall-runoff modelling. The study also highlights the usefulness of system-theoretic LRB models in simulating data of a temporal resolution that masks the high-frequency features of the processes being modelled.