Predicting The NFL Champion With Machine Learning
Using Machine Learning in preparation to predict the NFL Champion right before the playoffs for every up-coming year!
Article Orginally Made: 11/10/23
Article Updated: 11/17/2024 “Mid-Season 24–25 Predictions In!”
Introduction
The NFL playoffs are a 14 team single game elimination tournament, and one of the most exciting sporting events in the world. The Super Bowl itself has become an American tradition as it is the most watched sporting event every year in the United States. The complexity of the game, physicality, athleticism, and strategy is what makes American football one of the most popular sporting leagues.
Now in regards to the playoffs, many fans and analysts every year make predictions of why they believe this or that team has the best chance to win it all. As a sports fan, often times I find our feelings and intution can deceive us. Look the NFL playoffs is crazy, single-elimination format, not many regular season games, trades, I guess the question I am asking is…
“Can numbers even accurately measure the vast and wild NFL landscape…
Well that answer is…. Yes, kind of!”
I have came-back and spent more time analyzing the NFL post-season, identified some new trends, and managed to create a new model that has captured a better understanding of the NFL post-season. The model in particular has ensured that the eventual champion is with in the top 3 predicted teams…for the last 10 seasons.
NFL Season 24–25 Mid-Season Predictions
Year-by-Year Predictions | 1 for 2
Data
All the data used for this project comes from football reference. For this project I used player and team data that traces back to the 1991 season. For more information on the statistics click here for the glossary of the statistics.
Model Prediction Value
The value to be calculated by the model is champion_share. This is a custom statistic calculated from the ending post-season standing. An example of this calculation is the 2023 title runner-up Philadelphia Eagles who achieved 3 playoff wins out of the potential 4 wins it takes to win the title (3 / 4 = 0.75).
Something else to mention is that “bye” seeds 1 or 2, it depends on season, automatically get a playoff win in the calculation. This keeps it consisent amongst all other teams in the playoff field.
Machine Learning Model Results
Machine Learning Model Used: Random-Forest Regression (Complex)
For full details feel free to investiage: My GitHub
This section covers 3 charts that captures every aspect of what it takes to become a Modern NFL Champion.
- Season Playoff Comparison Chart (2024 Season)
- Feature Importance Chart (Most Important Statistics)
- Feature Importance Heat Table (2024 Season)
The Season Playoff Comparison Chart shows the differences between the prediction and actual results of the 2024 playoff.
One way to assess which features are most important for predicting the NBA playoffs is by examining the feature importance scores of the model. In this project, a complex tree-based model (XGBoost Regression) was used, which calculates feature importance based on how much each feature contributes to reducing error in the predictions.
However in this models case, the Feature Importance Chart doesn’t indicate whether a feature has a positive or negative effect — it simply shows how relevant each feature is for the model’s overall performance. This is a common trait of tree-based models, which split the data based on different features without assigning direct positive or negative weights to them.
Random-Forest builds multiple decision trees independently, with each tree trained on a random subset of the data and features. By averaging the predictions from all these trees, Random Forest produces a nice generalized and stable final prediction.
For more information on how tree-based models, click here.
Now, for a more detailed view, here is the Feature Importance Heat Table, which displays the top 10 most important features in a tabular format. This table not only highlights how these key features contribute to the prediction but also reveals how they relate to one another.
What Makes An NFL Champion?
With all of the information presented, this article can briefly be summarized in these 4 points (in ascending order on importance).
- The team should rank highly in my custom team rating metric, which assesses overall performance for the season. This rating considers the team’s regular season record, player star power, margin of victory, and kicker efficiency for that season. Click here for full details of the custom team season ratingstatistic(team_rating_custom, top_5_team_rating_custom, 40_50+_perc_x_makes).
- A strong offensive line and an experienced quarterback — ideally one who has received MVP votes or is a seasoned veteran — are essential. The quarterback should rank among the top seven in my custom quarterback metric, demonstrating efficiency, particularly in yards per completion, throughout the regular season. Click here for full details of the custom quarterback season rating statistic (passing_offense_Y/C, QBSRC_rk).
- A championship team also requires a strong, disciplined defense and offense (team_offense_Penalties_Pen), proven either in previous seasons or during the current season (team_defense_EXP). Importantly, the current season’s defense should excel at stopping the run game (team_defense_Rushing_Yds) and be equally adept at defending the pass (team_defense_Tot Yds & TO_Ply).
- Finally, the team must have a coach with extensive playoff experience. This includes coaching several past playoff games, which is crucial for navigating the high-pressure environment of the post-season (sum_coach_playoff_games).
Conclusion
Overall, out of all of my projects this was by far the most unpredictable when comparing it to my NBA and the Men’s NCAA Division I Tournament Champion projects. By leveraging a large dataset, creating numerous features to evaluate the strength of individual players and coaches, and lastly using a complex tree-based regressor. The model to my surprise delivered promising results despite the inherent unpredictability.
Feedback
Any questions or feedback are always welcome, so feel free to post them. For full details of all steps done for this project check out my GitHub.
