Spectral retrieval techniques are currently our best tool to interpret the observed exoplanet atmospheric data. Said techniques retrieve the optimal atmospheric components and parameters by identifying the best fit to an observed transmission/emission spectrum. Over the past decade, our understanding of remote worlds in our galaxy has flourished thanks to the use of increasingly sophisticated spectral retrieval techniques and the collective effort of the community working on exoplanet atmospheric models. A new generation of instruments in space and from the ground is expected to deliver higher quality data in the next decade, it is therefore paramount to upgrade current models and improve their reliability, completeness and numerical speed with which they can be run. In this paper, we address the issue of reliability of the results provided by retrieval models in the presence of systematics of unknown origin. More specifically, we demonstrate that if we fit directly individual light-curves at different wavelengths (L-retrieval), instead of fitting transit or eclipse depths, as it is currently done (S-retrieval), the results obtained are more robust against astrophysical and instrumental noise. This new approach is tested, in particular, when discrepant simulated observations from HST/WFC3 and Spitzer/IRAC are combined. We find that while S-retrievals converge to an incorrect solution without any warning, L-retrievals are able to identify potential discrepancies between the data-sets.