Association Rule Mining for Best Ball Transactions

Author: Bill Jones

Introduction

Once upon a time in the vast realm of fantasy football, a passionate adventurer set out on a thrilling quest. Armed with a deep love for the game and a fascination with data science, he embarked on an exhilarating journey to unlock the hidden secrets of best ball fantasy football. As he ventured deeper into the realm of best ball content, a dynamic data competition caught his attention: Best Ball Data Bowl. Intrigued by its potential, he delved into the depths of data hoping to unearth valuable patterns and insights that had remained concealed from the naked eye. Join me on this exhilarating journey as I attempt to the mysteries of best ball and rewrite the playbook for success. 

Thanks for bearing with me while I had a bit of fun with chatGPT, now back to talking normal…

BBM Basics

For those who may be less familiar with Best Ball Mania (BBM) by Underdog, it’s a unique and immensely popular fantasy football tournament. Best ball is a fantasy football format where participants draft a team and then have no in-season management requirements like waiver wire additions or setting lineups. Instead, the highest-scoring players are automatically placed in the starting lineup, and leagues are decided based on cumulative season points rather than head-to-head matchups.

BBM takes best ball to another level with its distinct two-stage structure: the regular season and playoffs. During the 14-week regular season, teams compete against others in their draft, with the top two teams advancing to the playoffs. The playoffs then consist of a series of single-week DFS-style competitions, where teams from different drafts, potentially with overlapping players to your roster, vie for advancement to the next round.

Data Source

Before I jump back into analytics, I would like to acknowledge the data source. The data used for this analysis was obtained from Underdog Fantasy and is the basis for the Best Ball Data Bowl. There have been three BBM’s but the Best Ball Data Bowl competition states the competitors should only be using the BBMII and BBMIII data as the BBMI data is a bit of a mess, so I’ll be following the same rules in my studies.

Association Rule Mining

Association rule mining is a machine learning method for discovering interactions between variables in a dataset of transactions and can be used to identify where, how, and why certain items may be connected. The classic use for association rule mining is grocery shopping analysis. To explain what association rule mining is here is an example.

A grocery store has 1,000 sales transactions in one day of varying product mix. Of those 1,000 transactions, 100 of them had tortilla chips and beer (Aeronaut Brewing’s A Year with Dr. Nandu is my current favorite for the hop heads out there – looking at you Matt Landes). Of those 100, 69 had salsa in the transaction as well. So, the rule for this situation would be…

{tortilla chips, beer} -> {salsa}

- Support – 69/1,000 = 0.069

- Confidence – 69/100 = 0.69

Support is how often the Left Hands Side (LHS) situation is happening across everything and then confidence is how often the Right Hands Side (RHS) occurs know the LHS had already occurred. 

There is a calculation called lift that I must mention if I am going to discuss association rule mining. Lift is confidence/support and helps with the analysis of if two events are independent of each other and therefore no rule could be drawn involving the events. This concept isn’t relevant to the later analysis so we’re ignoring it for this post. I would be happy to discuss if someone wanted to unpack this but I’m getting a little wordy and want to get to the substance. 

Connecting Association Rule Mining to BBM

So how does association rule mining work with the BBM data you may be thinking? It’s actually quite simple… Each team drafted is a transaction. In each transaction there is some combination of 18 players that were drafted together as well as the results of each team in terms of advancement to the next round that are used to create rules about the population. The best way to show this is with an example from BBMIII.

There were 12,287 teams that were able to draft the Lamar Jackson and Mark Andrews stack in BBMIII. Of those 12,287 teams only 2,081 advanced to the playoffs. This translates to an advance rate of 16.94% for teams that stacked up Jackson and Andrews. 

The traditional Support calculation would have the denominator as total transactions (count of total teams in the contest), but this doesn’t exactly make sense for the BBM tournaments as only one team in each draft room can draft a player (or combination of players), not every team in the contest and as such I looked for a better metric to describe the frequency of the rule. Based upon the BBM format I believe a more logical denominator would be the total number of draft rooms in the tournament, 37,600, which results in a Modified Support of 5.53% for the Jackson and Andrews stack. This metric is still not exactly presenting pick frequency like I was looking for, so I went with a simple Drafted/Stack Frequency metric which divides the total times the transaction occurred by total drafts. This stack had a Stack Frequency of 32.68%, which was the 5th most common pairing of all pairing from BBMIII. 

{Lamar Jackson, Mark Andrews} -> {Advanced to Playoffs}

- Support – 2,081/451,200 = 0.0046121

- Confidence (Advance Rate) – 2,081/12,287 = 0.169366

- Modified Support – 2,081/37,600 = 0.05534574

- Drafted/Stack Frequency – 12,287/37,600 = 0.32678191

Top Rules from BBMIII

With a solid understanding of association rule mining and its relevance to fantasy football, let's dive into some of the top rules from BBMIII. Here are the top, bottom, and most frequent rules for one, two, and three player combinations in the tournament.

1 Player Rules

2 Player Rules

3 Player Rules

It only takes two

I have a couple of thoughts on the figures above. First and foremost, the importance of player selection and picking the right players cannot be understated. With the base BBM advance rate being 16.67% (2 out of 12), it was clear that choosing the right players was key for advancing in BBM, as evidenced by players like Josh Jacobs and Tyreek Hill. Josh Jacobs had the highest advance rate, 46.17%, followed closely by Tyreek Hill, who showcases an advance rate of 35.73%. However, the real magic happened when one correctly identified and paired elite players together. Take, for example, the combination of Josh Jacobs and Tyreek Hill, this resulted in an advance rate of 73.36%. Moreover, it's worth noting that adding a third player, such as Patrick Mahomes, to the mix elevated the advance rate to a remarkable 90.5%. This highlights the true boost to a team’s value one can create by being right on just two to three picks. 

Additionally, it was interesting to observe that the most common rules in BBMIII didn’t always produce high advance rates. In fact, a closer analysis revealed that a significant number of these common pairings had below-standard advance rates, rather than surpassing the average rate. I believe this is a function of BBM format which has a DFS like tournament where week 17 stacks and correlations are needed to win the big money. I believe the format causes BBM entrants to make big bets on a single set of players which can all get torpedoed at the same time (e.g., Russell Wilson & the Broncos or Matt Stafford & the Rams). Additionally, I believe the format distorts the draft board away from traditional production-based drafting as popular stacks get steamed up the draft board, especially the second part of a stack attached to a top quarterback (e.g., Gabe Davis getting pushed up to round out the Josh Allen and Stefon Diggs combination). Who do I think the "Gabe Davis" is for 2023? Players such as Tyler Boyd (WR3 being pushed up by Chase/Higgins/Burrows drafters) or Zay Jones (WR3 being pushed up by Ridley/Kirk/Engram/Lawerence drafters) come to mind.

Conclusion

In conclusion, the figures above reaffirm the significance of player selection in BBM and how picking the right (or wrong) players can have a substantial impact on your team's performance and advance rate. By carefully considering and pairing players, you can increase your chances of achieving success in BBMIV. Which player(s) are you planting your flag on for BBMIV? I know Chris Spags is going all-in on Anthony Richardson.

Lastly, I want to extend my appreciation to all who have made it this far. I look forward to expanding my analysis on this dataset and potentially entering the Best Ball Data Bowl! Stay tuned for more captivating analytics posts as I continue my journey into football analytics and if you have any research ideas or suggestions, I encourage you to connect with me on Twitter (@NFL_Billy_Jones). 

Bonus - BBMII

*This blog post was enabled by ChatGPT. The text was generated by me, and the content is my own, but some sentences and wording were provided by the model. I take full responsibility for all information produced in this blog. More information about OpenAI and their technology can be found at https://openai.com.*