Recommendation systems help users find interesting items and reduce the information overflow problem on websites. Much research has been conducted on such systems the last decades, and there exist several recommendation approaches with different strengths and weaknesses. In this thesis, we investigate which recommendation approach or combination of approaches that can recommend the most interesting content for each individual user of the sports video application Forzify. We use previous research and literature to find the approaches that are best suited for the data gathered in Forzify and the features wanted in a new recommendation system. Three approaches turn out to be the best suited: item-based collaborative filtering, model-based collaborative filtering and content-based filtering. We implement one algorithm from each of these approaches and a baseline algorithm. The four algorithms are evaluated in an offline evaluation to find out which of the approaches that performs best in terms of recommendation accuracy, both for new and old users, and scalability. As Forzify so far has gathered limited user interaction data, we have to test the approaches on other datasets. To increase the validity, we investigate the accuracy of the algorithms in datasets from different domains. This makes it possible to check whether it is consistencies in the accuracy of the algorithms across the domains. From our evaluation of the accuracy of the algorithms, we can see both differences and similarities across the domains. The accuracy of the different algorithms is more even in some domains than in others, and some domains generally have higher accuracy, but there is a tendency that the algorithms performing well in one domain also do so in the other domains. Due to this cross-domain consistency, our results provide a good basis for choosing the best approach for Forzify. We conclude that the model-based collaborative filtering approach is the best choice for Forzify. It gives accurate recommendations for both new and old users across the datasets, and it scales well for larger numbers of users and items.