Game theory has been a buzzword in the DFS space since the beginning, and for most players, it boils down to simply finding a lower-owned leverage play to include in lineups. However, there are many more applications and practices that help you differentiate when you start to think about what Game Theory actually is and how it can be applied to overall lineup construction. Game theory is a branch of mathematics that studies how players make decisions in competitive games where the outcome depends on the actions of both parties. In other words, we are competing against other DFS contestants, making unknown actions, rather than trying to hit a target score. In DFS, game theory helps players think beyond just picking the "best" lineup and instead focus on how other players are likely to behave – this is why most flock toward ownership projections to try and be different.
Game Theory in DFS: Predicting Opponent Behavior
In DFS tournaments, your lineup isn't evaluated in isolation. You’re competing against thousands of others, and part of the strategy is predicting how they’ll build their teams. If everyone picks the most obvious, high-performing players, you’ll need to differentiate your lineup to have a shot at winning. This is where game theory comes in: it helps you weigh the risks and rewards of making contrarian picks, in order to gain a competitive edge.
To start, let's give a high level view of contest types and when this is most important.
Why Game Theory Isn’t (as) Critical in Cash Games
In cash games like 50/50s or head-to-heads, where roughly half of the participants win the same amount of money, game theory takes a back seat. The goal in these contests is simply to build a high-scoring lineup that’s likely to finish in the top 50%. There’s less need to worry about being different from the crowd. Playing the safest, highest-projected players, who are often highly owned, is usually the best approach. I am not much of a cash game player, so this can be tough for me. I often have to force myself to follow the crowd - it is usually better to play chalk even if you think there might be a slightly better play.
Why Game Theory Matters in Large GPPs
In large GPP tournaments where the prize structure is very top-heavy, finishing near the middle doesn’t get you much—so just playing the best-projected lineup won’t cut it. To have a shot at the big prizes, you need to separate from the crowd, and that’s where game theory shines.
In large GPPs, understanding ownership is still important. Players with high ownership are often safe but offer limited upside for differentiating your lineup. Lower-owned players can provide an opportunity to gain leverage on the field, and I still do set rules for most slates to force in lower owned player. However, ownership isn’t everything. Focusing solely on low-owned players without considering their actual chances of success is a recipe for disaster. It’s about finding the right balance between popular players and those with strong potential but lower ownership.
Target Score Equity
There are some great articles covering this, and tons of opinions - as a matter of fact there used to be a great article explaining a similar process that walked through the math of several players, but I can’t find the link anymore - if anyone can find that send it our way as I would love to link it here!
This is how I try to determine if a player’s range of outcomes are appropriate for their projected ownership - the goal is to calculate what I think a player should be owned compared to their projected ownership, positional scarcity, and overall player pool for the slate. Similar metrics are used in other articles, I refer to this as Target Score Equity:
- To start, I use our machine learning based point projections, and our simulation based variance to get a baseline for each player (in our projections this is simply the Projection and Standard Deviation).
- I then use our historical dataset to try and determine what a winning value might be for each player. To simplify, this might be 4*Salary in the NFL. This Target changes depending on the sport, position, and salary range for a player.
NOTE: This goes against what I said earlier, we are trying to beat opponents not hit a winning score! True, but since this data is aggregated over hundreds of slates I still like to use it as a starting point.
- Calculate the likelihood that a player can hit their Target score. This can be as simple as using a normal distribution to calculate the z-score for each player, and then using the cumulative distribution function to get a percentage. Again, depending on the sport and position I will use different distributions:
NFL tends to be gamma distributed
MLB Pitchers tend to be closer to normally distributed
MLB Hitters tend to be long-tail gamma distributed
NBA tends to be closer to normally distributed
- I then layer in positional scarcity - for each position, what is the total likelihood that players will hit their Target score. Maybe it is a terrible weather day in NFL with rain everywhere, and it looks like RBs will hit target everywhere but QBs won’t. In this scenario it becomes more important to find which QB is going to hit target compared to the others, knowing that a bunch of RBs will perform well compared to their position.
You then divide each players likelihood by their positional likelihood of hitting Target to get a measure of how much TSE each player has compared to their peers. Put another way, if RB1 only has a 10% chance of hitting target, but the overall likelihood for RBs is only 20%, then RB1 represent 50% of the TSE and sets up to be a good play.
NOTE: If your positional likelihood is much greater than overall positional ownership, it might be good to revisit the first 3 steps i.e. If we assume total ownership of RBs is 233% (2 RBs per team plus FLEX) but the total likelihood is 500%, then your numbers might be off along the way
- In practice, I often use TSE as my Exposure limits. I will try to use each player up to their TSE for my GPP player pool. For smaller contests I look at my player pool and compare this TSE to projected ownership and see where things are different. This helps me understand where my data shows there could be high leverage plays.
Where this gets complicated is understanding the correlations between players and how that impacts their TSE. If a QB has a high TSE but all of their teammates are extremely low, then you have to question the logic, unless we are assuming that QB will rush for 3 TDs. Our simulation process leverages our massive historical dataset to find and apply correlations across teammates and opponents. We then surface these simulation based metrics at the QB and Stack level. These simulated metrics, along with the calculated player TSE ultimately drive my player pool strategy.
Other Ways to Apply Game Theory in DFS
Outside of the typical ownership leverage angle, using Game Theory in roster construction is a must for large GPPs - The team over at One Week Season are the gold standard for this thinking in my opinion. In many ways Game Theory is just the ability to be different than other lineups in a contest. Ownership has a number tied to it so it is the easiest to talk to, but here are some additional ways I try to get different, some more obvious than others:
- Salary Ranges. the public will usually try and spend the entire allotted salary for a contest. Try a wider range, allowing your lineups to leave salary on the table. I will often use TSE to find a lower priced pivot that still could win a contest even if that lower priced pivot is higher owned. Think about it this way - total lineup ownership/product lineup ownership is important, but if two lineups have the same ownership level but one leave $2k in salary, that lineups is probably more rare
- Play with Stacking. Outside of NBA, stacking is usually a must. Using DraftKings as an example, in baseball the public typically trends toward 5 hitter stacks, so simply playing a 4 man stack might be good enough (it probably isn’t), but what about 3-3 stacks with the rest of the slots reserved for power hitters. Maybe roll with the typical 5 hitter stack, but only when it is up against a highly owned pitcher.
- Get comfortable with the math and build your own projections. If we take it literally, one way to “get different” is to be the only person in the world using your data. This might be tough depending on your skillset, but if you are using the same projections from the same big data providers as everyone else, without updating anything, then you might be gravitating towards the same plays as everyone else.
- Over and Anti-Stacking. Since the public is usually stacking NFL GPP lineups, I often like to include solo QB lineups when there is a good rushing QB matchup. Similarly, If TSE shows a lopsided matchup that isn’t reflected by Vegas lines, a 5 player game stack, or even 6 in certain situations can pay dividends.
- Leverage Simulations. Contest Sims can be a great way to get an understanding of the potential range of outcomes for your lineups. I use our sims heavily to down-select from my large lineup pool to my playable 150 GPP lineups. As I go through a sim run, I start to see the types of lineups that have high potential and work from them. This is why my DFS process is an iterative one. Analyze Simulated metrics, Build, Simulate, Iterate, Repeat.
- Understand the DFS Meta. NBA can be rocky to start the season as teams find their rotation. MLB can be hard to track as rookies are called up late in the season. NFL teams can be an injury away from their season becoming meaningless. As these types of changes happen in the real world, understand how your DFS strategy is impacted. For example, in MLB I tend to shift from 4 to 5 Hitter stacks later in the season when facing Pitchers who have thrown a bunch of innings.
Successful application of game theory in DFS is an ongoing process. The meta constantly evolves as players adapt, the season plays on, and new strategies emerge. Keep iterating to find your Edge.
