We introduce two new loss functions designed to directly optimize the statistical significance of the expected number of signal events when training neural networks and decision trees to classify events as signal or background. The loss functions are designed to directly maximize commonly used estimates of the statistical significance, s / s + b , and the so-called Asimov estimate, Za. We consider their use in a toy search for Supersymmetric particles with 30 fb−1 of 14 TeV data collected at the LHC. In the case that the search for this model is dominated by systematic uncertainties, it is found that the loss function based on Za can outperform the binary cross entropy in defining an optimal search region. The same approach is applied to a boosted decision tree by modifying the objective function used in gradient tree boosting.