Corruption is a major global problem. It undermines economic growth, government efficiency and public trust in institutions. Therefore, predicting the evolution of the corruption situation is a very important goal. However, choosing which indicators to use is difficult. The perception of corruption either measures a subjective corruption perception index or is based on objective statistical numbers. However, these latter statistics do not show undetected cases of corruption and therefore underestimate the frequency of the phenomenon. This study uses a machine learning method to select economic and socio-statistical indicators that can be used to estimate corruption. Based on the method, the trend of the relationship between the indicators included in the research and the perception of corruption or the lack thereof enriches the knowledge on the subject with a lot of information. The results also bring science closer to developing a reliable forecasting method.