Man fits linear discriminant classifier on small, noisy set of data, finds it has 100% accuracy on entire set. Did not do any cross-validation to figure out robustness of fitting procedure or magnitude of generalization error. News at 11.
To be fair he claims to have developed the system in 1980 and then used it subsequently to predict elections. But you're right that the data is noisy and small. And he did get it wrong in 2000 by predicting an Al Gore victory. So 7/8 correct on the "test set."