How do you explain the big performance boost for some of your benchmarks? The leaderboard scores are in the low 0.9 for Ashrae, yet you have unrealistic performance reported in here. Same goes for the classification challenge. A good data scientist should produce a model that is within 1–2% error from a good auto ml system. This tells me that you did something wrong during the modeling stage.

Can you provide more details about how you ran your benchmarks.

https://www.kaggle.com/c/ashrae-energy-prediction/leaderboard