This week, I am finally going to be able to get back to work that started a few weeks ago following a conversation I had when Kendrick visited. The main issue we are trying to tackle is the setting of regularization parameters when conducting a regularized regression. It turns out that at least in Ridge Regression (but maybe also in other methods? Lasso? Elastic Net?), you can treat this mathematically and derive something close to a closed-form solution that will tell you the effect of a particular value on the degree of regularization that occurs in practice. Thanks to the nice simulation methods available as part of Scikit Learn, it is relatively easy to do some simulations that could demonstrate this and test it in various regimes (e.g., correlated vs. uncorrelated regressors). I am excited about this project, because it really requires extending myself into terrain that I am not super confident in, and would like to gain some more confidence in. There’s another statistics paper that I still need to write, once this one is done.