This study aims to describe pre-reproductive movements of Luciobarbus bocagei and Pseudochondrostoma duriense in a regulated canyon-stretch of the Tormes River (Spain), with high environmental variability, semipermeable barriers, and fishway retrofitting actions. The main objectives were to identify peak migration dates and environmental drivers, test ensemble-learning techniques to model fish migration and propose adaptive management measures. To achieve this, fish movements were 5-year monitored in a stepped fishway and Survival Analysis and Random Forest techniques were used for data analysis and modelling. Results showed that migration occurred in May–July, a wider period than the one previously reported in the literature. Movements were triggered by the increase in water temperature and photoperiod, and were strongly affected by the hydraulic river scenario (water levels and discharge) at the semipermeable barriers. Random Forest was able to include the effect of each barrier and predict accurately timing and number of migrants, classifying and ranking the importance of variables. Moreover, developed models allowed to assess fishway retrofitting actions and to predict positive effects in fish number under new, scheduled and variable environmental flow scenarios. Long-term monitoring together with ensemble-learning methods can allow the definition of cost-effective adaptive management strategies to ensure endemic fish conservation.