To compensate for no-shows and cancellations on short notice, air cargo airlines sell more than their physical capacity. This revenue management practise is called overbooking. It originates from the passenger aviation business for which a wealth of overbooking research already exists. Due to a higher complexity in the air cargo business, however, this passenger research cannot simply be transferred to the air freight sector. We complement the relatively sparse air cargo overbooking research by estimating the no-show respectively cancellation probabilities of individual bookings, from which overbooking levels can then be calculated. By doing so, we abandon the traditional approach of estimating optimal overbooking levels based on the flight as a whole. The estimations are obtained through machine learning. More specifically, we use random forests as they dominate over other classifiers. Overall, the resulting models perform well although we are able to predict no-shows more reliably than cancellations. To take a step towards transforming these estimates into overbooking levels, we provide application tests and ideas on how to proceed.