18 Million Brackets Later, Here Are the Best — and Worst — Ways to Fill Out Your March Madness Bracket

Ben Lahner
3 min readFeb 18, 2023

--

Silhouettes of five basketball players
Generated by Stable Diffusion using the prompt, “​​people picking basketball teams playground silhouette high resolution 4k detailed winner”.

Every spring, tens of millions of people fill out a March Madness basketball bracket in hopes of being one of the select few to win mind-boggling cash prizes and luxurious vacations. While predicting the winners of only 63 games may sound simple, a winning bracket must strike a delicate balance between guessing the cinderella-story upsets and letting the tournament’s favorites dominate.

The million dollar question is, quite literally, what is the best way to fill out a winning bracket to have a shot at the big prizes? I tested 18 million brackets over the past 6 tournaments using 3 popular bracket prediction methods to find out.

The first method I tested, “High Seed”, always chooses the better (higher) of the two seeds to win each game (Figure 1, top gray). Any reasonable bracket recognizes that higher seeded teams are more likely to win games — and the tournament — than lower seeded teams.

The second method, “Coin Toss”, decides each game with a 50/50 coin toss, regardless of team seed (Figure 1, middle brown). This method is your best bet to predict those stunning upsets that shatter brackets every year.

The last method, “Sportsball”, uses the Sportsball Bracket generator to probabilistically decide each game based on the teams’ seeds (Figure 1, bottom yellow). For example, if a 4 seed plays a 13 seed, the 4 seed team will have a 13/(4+13) = 76.5% chance of winning*. This method strikes a necessary balance between the “High Seed” and “Coin Toss” methods by giving a clear advantage to high seeded teams while allowing for upsets.

Figure 1 shows the backtested performance of 1 million brackets per method over the last six years (18 million brackets total). The green bar depicts the Big Money zone, the range of scores that won the millions of dollars in prizes in the nation’s largest bracket challenge competitions — ESPN, NCAA.com, FoxSports, CBS, and Yahoo. For reference, the gold bar on the far right shows the 192 point score needed for the elusive perfect bracket (using the standard scoring system of 1, 2, 4, 8, 16, 32 for rounds 1–6, respectively).

The “Sportsball” method consistently outputs unique and high-performing brackets each year, clearly outperforming the “Coin Toss” and “High Seed” methods. It is the only method that regularly generates brackets in the Big Money zone, and, in 2017 it would have even generated the nation’s best bracket. Knock on wood, “Sportsball” brackets come within just a few games of a perfect bracket each year.

Compared to the “Coin Toss” method, “Sportsball” brackets are better in every aspect — higher minimum score, average, median, maximum score, and number of brackets in the Big Money zone. While the “High Seed” method offers an attractive minimum score, the historical backtests show that it is the least likely method to generate a bracket in the Big Money zone. If, however, the “High Seed” method does generate a winning bracket, it still has a fatal flaw — it can only generate eight unique brackets (brackets are identical through the all 1-seed Final Four), meaning you will have to split any winnings between the many thousands of other people using this method. In contrast, the “Sportsball” method regularly generates unique, top-performing brackets each year for both the men’s and women’s tournaments, making it your best bet to strike it rich this March.

Violin plots analyzing three bracket prediction methods over the past 6 March Madness tournaments
Figure 1: For each of the previous six years of the men’s NCAA March Madness tournament, we generated 1 million brackets from three methods (3 million brackets in total per year) and evaluated them against that year’s ground truth bracket. We score correct game predictions using the standard 1–2–4–8–16–32 points per game system in rounds 1, 2, 3, 4, 5, and 6, respectively. Violin plots show the distribution of bracket scores with ends extending to the lowest and highest observation. The white circle in each violin plot depicts the median score. The green vertical bar depicts the spread of bracket scores that achieved first place in the largest bracket challenge competitions. The gold bar at Bracket Score 192 shows the score needed for a perfect bracket. Note that there was no NCAA March Madness tournament in the year 2020.

*”Sportsball” artificially deflates the 16 seed to a “30” seed, meaning a 16 seed will have a 1/(30+1) = 3.2% chance of beating a 1 seed.

--

--

Ben Lahner

I enjoy tackling questions that appear impossible to answer. Current PhD student @MIT. Website: https://blahner.github.io/