Road traffic accidents pose significant problems on Swedish roads, not only in their own dev- astating nature for the people involved but also because they tend to disrupt traffic flow. The Swedish Transport Administration, Trafikverket, actively works towards decreasing the negative consequences of accidents through coordinating traffic and communicating information to SOS, road assistance and other personnel in order to efficiently handle the incidents. Road traffic op- erators in Stockholm and Skåne publish announcements of accidents, in which they describe the type and location of accidents as well as give an end-time estimate, for when they think the state of the road and traffic flow will return to normal operation. The accident data is registered into a database called Nationellt Trafikledningssystem, NTS, where an array of information is stored for each accident, together with mentioned end-time estimate. This thesis explores the possibility of using predictive analysis, machine learning, in the task of predicting traffic accident end-times, based on historical road traffic accident data. The goal is to make estimates that outperform the manual ones set by the road traffic operators and hopefully prove that such techniques can be used in the future. Through preprocessing, NTS and meterological data were merged and transformed for a wide range of machine learning models with the top performers being Extreme Gradient Boosting, XGBoost and Support Vector Regressor, SVR. The thesis concludes that it is possible to successfully use machine learning models to predict end-times for accidents on the roads of Stockholm and Skåne, while also outperforming the ones set by Trafikverket. It is however important to consider the complex system of road traffic. Some accidents’ end-times are more important to predict than others depending on time, place and severity. Suggested im- provements include the use of more detailed attributes, more descriptive of the actual accidents, as well as better quality control for NTS data registration. The results of this thesis can be seen as a proof of concept and an assistive tool rather than an applicable method.