making your own college football betting lines

I tested the accuracy of this model in two ways–one being the generic, “what percent of the model’s” bets were correct (which certainly is helpful), and the other using F-Scores (which improve accuracy judgments on imbalanced datasets). It seems more unusual to see away penalty yards, away return yards, and away passes as the other top three influencers, but you really never know what metrics are going to help your model gain an edge. It makes a lot of sense that the Vegas spread is the most influential for this model–after all, what better way to spread lines than by using the spread that the market provides? Home yards being influential also makes sense, as the spread is usually set to the home team, which means their yard total would have a high influence on what our model predicts. In the graphic below, we can see which metrics affected each prediction the most, according to their SHAP values (each blue point shows a data value for a single feature and how widely it affected the model prediction–SHAP value). In the case of ALICE, we are trying to predict the spread of NCAAF games using the 5-game rolling mean metrics of the following: Vegas spread, yards, touchdowns, penalty yards, return yards, kick yards, turnovers, takeaways, sacks, EPA, rushes, passes, avg. Without further adieu, let’s jump into this article and learn how ALICE runs.Īs we know, extreme gradient boosted models are predictive models that intake a number of metrics in order to make each assumption. My REBEL model performed excellently on its tests for the NCAAF and the NFL, making me eager to test the limits of the boosting technique. Another boosted model? What is it with this guy? While this is my third time using this method, there is a saying that states ‘don’t fix what ain’t broke’, which seems to apply to my situation. I understand that this article may be a bit lengthy, so if you want to skip around a bit, here are the links to each section In short, it is a supervised learning algorithm, which attempts to predict target variables by combining prior predictions from weaker and underdeveloped models. For those of you who aren’t familiar with ‘boosted’ models, you can learn more about them at this link. This model is built using an extreme gradient boosted algorithm, similar to REBEL’s. If you haven’t read my articles about my first betting model, REBEL, the links are here and here for your convenience.

In this article, I will be breaking down my newest betting model, ALICE, which stands for Adaptive (spread) Learning (with xGBoost) Corrective Error (techniques).

Making your own college football betting lines