R2 Ted Fantasy Football is a a sophisticated projection system for predicting weekly fantasy football outcomes, with a focus on daily fantasy football. It utilizes cutting edge machine learning linear models and a large data set including advanced stats, climate, travel distance, playing surface, stadium details, days rest, matchup-specific statistics, and more to predict each statistical category that accrues fantasy points. R2 then aggregates fantasy point estimates using those categorical projections, constructs standard deviation estimates for each player, and simulates the fantasy week 10,000 times. Using those simulations, we can calculate the probability that each player appears in the optimal lineup that week. R2 Ted is new for the 2022-23 NFL Season and should be considered a beta version that we're testing.
At R2, we believe you can simplify success in large DFS tournaments to three key elements:
1. Sharp projections - You want the sharpest prediction system available to establish the expected or average level of performance given all information available and quantifiable.
2. Prediction interval - You need a reliable method to approximate the distribution of potential fantasy points a player could score, given their projected score. In other words, you need to be able to approximate each player's ceiling and floor. This gets tricky because some players' performances are more volatile than others week-to-week. For instance, a receiver that accrues fantasy points largely via receptions and chewing up yards may be more consistent than one that gets most of their points from long TD receptions. As such, you need a prediction interval that is unique to each player.
3. Research - R2 Ted accounts for a lot of factors that could impact a player's fantasy total. In fact, if you can think of it and it can be quantified reliably, then R2's raw projections likely account for it as best as it can. That said, R2 Ted can not account for everything. For instance, the raw model won't know which players are likely to see increased playing time due to injuries or suspensions, won't know if the best linebacker on their opponent is out on paternity leave, if there has been a last minute change in the weather situation, etc. However, R2 will give us the tools to deal with many of these scenarios. More on that below.
First, let's focus on how R2 Ted's raw model performs on projection and prediction intervals. Chart 1 shows how raw R2 Ted projections perform (there are no post-estimation adjustments based on research) vs. the actual fantasy totals. Chart 2 is an example of how R2 simulates fantasy performance based on its raw projections and prediction intervals. Comparing these two charts allows us to gauge how well our model will do when we simulate the week with some random error to create DFS lineups.
Chart 1 - Actual Fantasy Points vs. Raw Ted Projected Points
Chart 2 - Simulated Fantasy Points vs. Raw Ted Projected Points
Before digging in too much, it is worth pausing to reflect on a few things about the data. First, fantasy football scores are noisy - there is a very large amount of variance in a player's output week-to-week. Second, the noise in the data increases with the level of the R2 projection - a statistical property known as "heteroscedasticity." This means that volatility in player performance increases as the R2 projection gets higher. Third, superstar players that average 20+ points really are significant outliers in our data set.
So how does our model do in matching the actual distribution of fantasy scores given our underlying projections? Overall, I think it does pretty well. The errors are generally distributed in our simulations as they are in actuality, demonstrating higher volatility the larger the R2 Ted projection. However, if you're a seasoned DFS vet, there is likely a region of these charts that is already screaming out at you. Let's show it in Chart 3.
Chart 3 - The Research Miss zone
If you compare the actual performance to simulated performance charts, you'll see a significantly higher frequency of 20 and 30 point games amongst players projected to score between 2-12 points in the actual vs. projected chart than the simulated vs. projected chart. These would be extremely costly misses in a DFS roster - cheap players scoring huge points! These should be exceptionally rare events given the projection and prediction intervals in that range - and in many cases, they should be astronomically rare. So rare, in fact, that we can safely conclude that these are simply research misses - or cases where the model didn't have enough information to make a reliable forecast - such as backups moving up the roster to fill in for an injured player (particularly backup QBs and RBs).
The good news is that we've actually designed Ted with this research miss zone problem in mind. As mentioned above, R2 Ted projects each scoring category independently. It also projects variables that give players the opportunity to score fantasy points - such as pass attempts, carries, snaps, targets, and more. As we go about our routine research each week, we will be able to scale players' projections and prediction intervals up and down as appropriate. R2 will be able match its current team depth charts with injury data as it is released, and then ensure that replacement players seeing expanded roles are projected to get the number of opportunities a player in that role should get.
In DFS you want to take advantage of correlations between players to give your roster a higher ceiling. A primary method to achieve this is to stack a quarterback in the same lineup with one or more of his receivers. To estimate the probability that a QB and WR/TE tandem appear in the optimal lineup that week, we simulate the week with some correlation between QB performance and WR performance using a coefficient we estimated empirically. This allows us to calculate the probability that a combination of a QB and one of his receivers appear in the optimal lineup that week.
So there you have it - R2 NFL Ted will provide all three elements of a complete and successful projection system. It will use cutting edge machine learning techniques to obtain fantasy point projections and the player-specific prediction intervals. It will also provide the research angle by maintaining up-to-date depth charts and matching it with injury data scraped from the web (as well as following the news). Will we be successful? Stick around and find out! If you achieve some success using our system - please let us know via this website or on Twitter (@DawgQuant).
A final note - at this time R2 does not yet have the plumbing to forecast rookies. For the first three weeks of the season, our rookie predictions will be based on outside research. By the start of the 2023 season, we will have developed a system to forecast rookie fantasy points.
How R2 Ted will be Updated
I will publish R2 Ted projections at least twice weekly. The first publication will be mid-week to start honing in on your target plays. The second publication will come after the Friday injury reports each week (either Friday or Saturday), which will reflect the best information to date on who is likely to play that week.
There are always last minute changes to the projections when the last injury news hits an hour before kick. I will not have time to update the site and play, but I will agree to email it to those that subscribe.
I will also provide updates via twitter @DawgQuant.
How to use R2 Ted
Currently, R2 Ted is only compatible with Draftkings contests. Ted was built to simulate each week and create large numbers of lineups to enter into large DK tournaments. There are many excel and web-based optimizers out there that are able to take R2 Ted projections and build lineups. I personally do this process in the R statistical package and will release an optimizer custom built for R2 projections and prediction intervals. To use our projections with outside optimizers, we recommend you download the projections in .csv format from R2 Ted Fantasy Football.
You may also use R2 Ted to build cash lineups. In this case, you want to use the central projection and avoid players with low floors and high standard deviations. Finally, you can use R2 Ted make season-long roster selection choices and inform prop bet decisions (though note that unlike how R2 Roy has been tested extensively ATS, R2 Ted has not been tested in prop bet markets).
What's coming for R2 Ted
Sample model output from 2021, Week 18 is currently available at R2 Ted Fantasy Football. When all information is available run projections for week 1, we will update the site and notify via twitter.
Things to be on the lookout for include a R2 optimizer in the R programming language with plenty of instructions custom-built for using the R2 Ted Fantasy Projection System, and a system for predicting fantasy points for kickers, which would make R2 compatible with Fanduel competitions.