Live Analysis of DraftKings' NBA Moneyline Bets with Machine Learning

NOTE: I live in California now, where online sports betting is illegal, so this project/site is no longer active. The tables displayed are from previous moments in time when I lived in Pennsylvania. Here is a link to the GitHub repository containing the code: GitHub Repository.

Welcome to BC Bets! This website uses machine learning to identify favorable NBA money line bets. It pulls all pregame lines listed on DraftKings' website every hour and evaluates them. For each game, it picks the better of the two lines and calculates a “confidence”, with higher confidence corresponding to a better bet. The goal of course is to make money!

Doesn't the House Always Win?

The central idea behind this methodology is that sportsbooks do not necessarily set moneylines based on how likely each team is to win, but rather based on how much bettors have placed on either team. This means that due to the large number of people who simply place bets on teams they like or on intuition (“this one's a lock bro I swear”), a good predictive model can be used to give consistent positive long-term returns.

Creating a Model

The first step in this process is creating a model that can predict the win percentages for two teams that will face each other based on their stats. The process for creating the model is broken into the following steps:

1. Data Collection: Statistics from every modern NBA game are available publicly. Data from the past 10 years was used to train the model.
2. Feature Engineering: This critical step identifies/computes stats that are used to train the model. After testing several combinations and reviewing existing models, the following three stats were used:
   1. Team ELO: A rating system that reflects a team's win percentage adjusted by the strength of schedule.
   2. Plus-Minus over the past 10 games: Captures recent performance trends as teams improve or worsen over the season.
   3. Home/Away Team: Since the home team wins 63% of the time, this factor is key and fully independent of other stats.
3. Model Selection: The method requires predicting win probabilities rather than just winners and losers. After testing various models, a linear regression model performed best.

The model trained on these input stats predicts NBA game outcomes correctly 68% of the time. This may sound low, but underdogs generally do win roughly 30% of the time, and the best currently existing model I could find by FiveThirtyEight is correct 70% of the time. And again, the main advantage of this model is not that it predicts a binary game outcome, but that it gives a quantifiable win percentage for each team. You may be able to predict game winners well, but how do you know if +340 or -500 is a better bet for a particular game? Or if they're both bad bets? That's what really matters in sports betting at the end of the day.

Identifying Good Bets

Okay now we have a model, but how exactly is it used? The key step in assessing the quality of a moneyline is computing what a “fair” win percentage is for a given moneyline is. For example, if a team has a 50% chance to win, +100 is a fair line (you will exactly break even in the long term). If a team has a 33% chance to win, then +150 is a fair line. You can compute a fair win percentage for any given moneyline. Once you have this, you simply compare the sportsbook's fair win percentage to the win percentage predicted by the model. If the sportsbook's fair win percentage is 50%, and the model gives that team a win percentage of 55%, then this is a good bet with a confidence of 5%. This site actively pulls all moneyline bets from DraftKings' website (updated every hour, pregame lines only) and goes through this evaluation process on each individual bet. For each game, it picks the better bet (higher confidence), and displays that along with the confidence in the “Predictions” tab of this site. Note that some lines will have negative confidence, which simply means both bets are bad, and it just picked the less negative one.

And that's it! I would use the above table to place my bets.

Weaknesses

In reality, you cannot simply rely on all green bets because there are events that the model cannot immediately account for. Since the features take time to recalibrate, factors such as player injuries, key players sitting out at the end of a season, or recent trades can drastically alter a team's performance before the model has a chance to adjust.

So, some level of active management is still required—follow the news and be cautious of green bets where a major change has occurred in the past few games. Allow the model time (typically over the next 5-10 games) to incorporate these changes.