The ability to make realistic judgments of one's performance is a demonstration of the possession of strong metacognitive skills. In this study we investigate the relationship between accuracy of self-evaluation as an expression of metacognitive skill, and learning gain in stoichiometry. The context is an academic development programme at a South African University offered for under-prepared students enrolled for science and engineering. These students generally exhibit unrealistically high levels of confidence in performance and this could potentially place them at risk by negatively affecting decisions regarding time management and self-regulation. We investigated whether overconfidence before instruction is corrected upon exposure to teaching. A three-tier stoichiometry test was used to collect qualitative and quantitative data before and after instruction. Findings indicate that the majority of the students were overconfident in the evaluation of their performance in both the pre- and posttests. Overconfidence was not a debilitating disposition when demonstrated in the pretest provided that it was corrected during teaching and learning. The most vulnerable students were those that judged their performance or lack thereof realistically in the pretest but became overconfident during the teaching and learning of stoichiometry. Our results suggest that under-prepared students are slow in developing accurate metacognitive monitoring skills within a classroom environment that did not include instruction focused on the development of such skills. We recommend a proactive and constructive response by educators which may reduce the incidence of failure and preserve the positive contribution of confidence, albeit excessively positive.