We can use N-fold cross-validation to do hacky regularization. sklearn has a lot of nice tools for thus. More or less, you can use a GridSearchCV object to do this. How it works is that you split the data into N sections. You loop and leave one of them out, fit a model, then predict for the other one. At the very end, you combine all of the out-of-sample predictions.

With this technique, we can loop through a range of regularization amplitudes, run CV for each of them, and pick the one that has the minimum chi2 or w/e.

It fits a lot more models, but would do the trick.