Text this: Response versus gradient boosting trees, GLMs and neural networks under Tweedie loss and log-link.