Alternatively the can do back testing - using historical data they feed a subset into the predictor, then compare it's predictions to actual history