Marketing mix models are tools used by marketers to understand how the different marketing mix levers generate sales. They are also used to make prediction of future sales given a marketing plan. These goals are often counterposed. Indeed, different model specifications can be used to achieve a greater explanatory power of past data or more accurate prediction of future ones. The aim of this thesis will be to understand whether different modelling approaches (i.e. neural network, random forest, or others machine learning algorithms) can provide more accurate prediction when compared with traditional marketing mix models.