tl;dr: Cheat sheet for linear regression metrics, and common approaches to improving metrics
One of the many reasons we care about model evaluation. Image courtesy the fantastic XKCD
I'll cut to the chase; linear regression is very well studied, and there are many, many metrics and model statistics to keep track of. Frustratingly, I've never found a convenient reference sheet for these metrics. So, I wrote a cheat sheet, and have iterated on it with considerable community input, as part of my role teaching data science to companies and individuals at Metis.
I'll also highlight that most of my work has been in leading Deep Learning and fraud, which have rarely involved linear models; I am, by no means, a domain expert. I've used this point of view to help write this reference for a general audience.
Below are the most common and most fundamental metrics for linear regression (OLS) models. This list is a work in progress, so feel free to reach out with any corrections, or stronger descriptions.
The natural next question is "What happens when your metrics aren't where you'd like them to be?" Well, hen, the hunt is afoot!
While model building is more of an art than a science, below are a few helpful (priority ordered) approaches to improving models.
- Trying another algorithm
- Using regularization (lasso, ridge or elasticnet)
- Changing functional forms for each feature (e.g. log scale, inverse scale)
- Adding polynomial terms
- Including other features
- Using more data (bigger training set)