logLoss
Binary log loss (cross-entropy) -(y ln(p) + (1 - y) ln(1 - p)) for predicted probability p and outcome y. p is clamped into [1e-15, 1 - 1e-15] so a confident wrong prediction scores a large finite loss rather than infinity.
maeLoss
Mean absolute error |prediction - truth|; less sensitive to outliers than mseLoss.
mseLoss
Mean squared error (prediction - truth)^2. For probability forecasts against outcomes in {0, 1} this is the Brier score.
pinballLoss
Pinball (quantile) loss at level tau: max(tau (y - yhat), (tau - 1)(y - yhat)), which is half the absolute error at tau = 0.5. Scores a model that predicts the tau quantile.