in contrast to parametric, semiparametric, and nonparametric approaches to inference, a hackometric approach completely lacks theory. it is any approach which essentially throws a bunch of machine learning algorithms at a problem and gets some kind of solution.
our nonparametric classifier estimator worked ok. our hackometric approach works as follows: log-transform the data, apply SVD, keep the first 6 singular values & vectors, and the last 2, and then put them through a SVM with a RBF kernel. this improves results by 0.002%, which is significant at alpha=0.05 level (uncorrected for multiple tests).
February 04, 2012