While I agree with your general statement, my gut says a well parameterized/regularized deep learning solution would perform as well as an ensemble of statistical approaches (without the expertise needed to select the statistical approaches) but would be harder to explain/interpret.