One of the many human capabilities acquired during adolescence is the adaptivity in changing environments. In this longitudinal study, we investigated this adaptivity, as measured by probabilistic reversal learning (PReL) tasks, in N = 143 adolescents at ages 14, 16 and 18. Computational modelling and functional magnetic resonance imaging were applied to identify the neurocognitive processes underlying reversal learning and its development. Previous studies have demonstrated a correlation between heavy alcohol use and impaired reversal learning. Our hypothesis was that PReL is negatively associated with current and future alcohol use and that alcohol use impairs PReL by altering neurocognitive processes. Behaviourally, PReL performance improved, which was associated with a lower probability of switching choices and was considered an adaptive process. Computationally, this was accounted for by higher learning rates, enhanced sensitivity to wins and reduced sensitivity to losses in older adolescents. Alcohol consumption increased but remained at a low level for most participants. More risky drinking was associated with less medial frontal activity elicited by reward prediction errors. These findings suggest that reversal learning may be more relevant for the maintenance or escalation of risky than for low-level drinking. Challenges and potential solutions for longitudinal studies such as reliability are discussed.