The use of socio-technical data to predict elections is a growing research area. We argue that election prediction research suffers from under-specified theoretical models that do not properly distinguish between 'poll-like' and 'prediction market-like' mechanisms understand findings.