Peer-to-peer ridesharing (P2P-RS) enables people to arrange one-time rides with their own private cars, without the involvement of professional drivers. It is a prominent collective intelligence application producing significant benefits both for individuals (reduced costs) and for the entire community (reduced pollution and traffic), as we showed in a recent publication where we proposed an online approximate solution algorithm for large-scale P2P-RS. In this paper we tackle the fundamental question of assessing the benefit of predicting ridesharing requests in the context of P2P-RS optimisation. Results on a public real-world show that, by employing a perfect predictor, the total reward can be improved by 5.27% with a forecast horizon of 1 minute. On the other hand, a vanilla long short-term memory neural network cannot improve upon a baseline predictor that simply replicates the previous day's requests, whilst achieving an almost-double accuracy.