The error on the test set is an approximation of the true future error. How close is it? We show how to compute a confidence interval [a,b] such that the error of our classifier in the future is between a and b (with high probability, and under the assumption that future data looks similar to our test set).