Did some on our internal datasets. It did not beat our AutoARIMA baselines, they are pretty noisy series though. Custom ML models only do a little better.