I suspect this is as much about profiling users (who will normally predict the outcome they subconsciously favour) than it is about effective predictions.