I've been working on this Kaggle competition a few weeks and I haven't been able to break the 0.40 RMSLE barrier. I've used more or less the same ideas described in this post: the cas/reg prediction split, extracting month/date/year data, etc. I feel like I'm missing a very helpful non-obvious feature that the better models(top 100) must have stumbled on. Any ideas on what this might be? Or perhaps it's as simple as created separate models for more specific cases of ridership like (cas+working day), (cas+nonworking day), etc.
