By using a custom loss function that penalizes low predictions more than high predictions, you can coerce the network to make high predictions to a specified quantile value, such as 90th percentile.