My scripts for modeling the bike sharing demand data from Kaggle. For more info: https://www.kaggle.com/c/bike-sharing-demand

Explore the data by installing the latest ipython (and pandas, matplotlib, numpy) and typing

ipython notebook

You can plot the rider count vs many input variables, while restricting the plot to arbitrary slices of input space.

This is my custom brute-force optimization stuff. No scikit-learn here. My models combine parameters and inputs in various nonlinear ways and just send the result into scipy optimization routines.

I spent a little time developing a framework that helps the model-builder return the gradient of a composite model function using sum and product rules for derivatives. This helps convergence in the scipy optimization routines.

See models.py and several model implementations in nonlinear.py.

This uses scikit-learn's support vector regression (SVR) module. As a black box, it does not perform great. I had to mess with the kernel to get decent approximations. Highest score on Kaggle was about 0.6, which is pretty far down the leaderboard.

MIT