INDUSTRY FORUM

Comments

  • HopscotchAnon

    Hey all,

    I am a high-volume DFS professional, and wanted to try and give back a bit to the community by trying to answer questions from a different perspective. I think the more transparency there is in the industry, and especially from the sites and professionals, the greater confidence the general public and other participants will have in the industry. And more people playing DFS is better for everyone involved – the casual players, the major sites, the sites built around DFS (such as RotoGrinders), as well as myself and other pros.

    Some background: My background is in STEM. I started playing DFS later than most, well after it became well known, and started with a typically small bankroll. As my models succeeded, my bankroll has scaled to the point that I now enter nearly all available tournaments in the sports I play. I play multiple sports, and all the major sites. My career entry fees total in the eight figures, and my trailing 12-month ROI is comfortably double-digits. This username is not my playing username.

    Some thoughts/observations on strategy and the industry to get the discussion rolling (and my views are my own, I certainly do not speak for anyone else):

    1. There are many ways to generate profitable models. Generally, more data is better, as long as it is properly handled. I personally don’t watch the sports I play – my models are entirely data-driven and automated, outside of interactions with the sites themselves. I do believe that knowledge of the sports and manual adjustments can improve models, and many pros benefit from such adjustments, but manual adjustments are difficult to backtest.

    2. Publicly available projections (paid or free) run the gamut from very little predictive power to very good. It is important to be able to know which projections actually work. My models incorporate both my own projections, and projections from a few specific sources. They help to account for extraneous shocks to specific games that my models can’t fully account for based on collected data alone.

    3. MME tournaments are in no way a guarantee of profit. MME is typically a multiplier of your edge, and if you have a negative edge, it will magnify your losses. There are a staggering number of possible lineups, even on small slates. The average overlap among pros on main MME NBA slates is only about 6%, so any two pros will typically have less than 10 lineups in common on any given slate.

    4. I understand many people play DFS for entertainment, and can get frustrated by professionals. But do understand that wherever there is significant money involved, there will be opportunity for those who wish to capitalize on it. I am very much in favor of the restrictions sites impose on the pros. For example, I can’t enter any tournament with an entry fee less than $3, and only large tournaments where fees are less than $5. This applies across all sports, so if I decided to play a new sport, I would have to enter higher dollar tournaments from the very beginning. For those playing for entertainment, the lower fee contests are a fantastic option.

    5. I won’t comment on specific things, but most of the conspiratorial ideas about pros and the sites don’t seem to based in a lot of fact, at least nowadays. The sites are friendly, and encourage the pros, but (at least for myself) they don’t provide any special favors. They are heavily motivated to try and keep the games clean and user’s confident in them, although I do agree some of their processes can be made more transparent.

    6. The well-known names have widely varying ROIs. Professionals tend to average a trailing 12-month ROI in the upper single digits, but with lots of variance, and there are also some that are negative (especially now). ROIs have tended to drop as time goes on, especially after the initial years of DFS. Competition adapts and improves, and some pros seem unable to. Rankings are not a great indicator of success, as they generally show volume, and have little correlation with ROI and profit.

    Happy to answer any questions!

  • HopscotchAnon

    @BrianVT said...

    Do you prefer R, Python, or other?

    Do you play PGA? If so, care to share whether you find regression-style techniques or classification (clustering golfers) to be better predictors? I’ve started leaning heavier on classification lately.

    I prefer Python over R, but I use a few other languages as well depending on the task.

    I use some form of classification in all the sports, some derived from clustering, but I also find it easy to go too deep to the point where the models get overly complex with little corresponding improvement. Regression-style techniques are used heavily throughout. Just like pruning neural networks, where you can reduce the complexity by removing nodes while keeping similar performance, there are many parts and features of the models that get too complex and at some point I reduce them back down to a much simpler, but similarly effective, model.

  • HopscotchAnon

    @dumbo said...

    Wouldn’t you agree that there is a critical mass of the number of optimal entries in mme (or similar entry contests) in re peak roi?

    If so, what is the number?

    I would say it’s about 15% of mme and scaled down at other entries. I never mme a contest and max out at 1/8 allowed entries and usually far less.

    The reason being, if you are contrarian, you’re 1sts are going to be the only prize you play for and after you’ve won that, the remainder of you’re entries are negative expectation.

    Thoughts?

    Generally, no, there isn’t a critical mass of optimal entries. Every lineup I enter (even the 150th in a 150-max contest) has positive expectation going in. But, sure, given one of my lineups wins, the conditional expectation of some of my other lineups will likely be negative. Unfortunately, I have yet to figure out how to predict when I will win! If the entry limits were much higher, perhaps in the thousands, then there may arise a critical mass of optimal entries.

    Regarding volatility, yes, it is important. I don’t think anything trumps raw expectation, but any significant parameter that can more accurately describe what you are trying to predict (and the relationships between them) is important.

  • Pidinolo

    Ok, a few of these have been touched on, but I have a mildly different approach. I have found that most of my success has come in single entries $ 100 + due to the fact I feel like a larger probability of playing the field gives me an edge. I tend to approach those with a more GTO consideration vs lower mme buyins. I find the “pros” tend to be fairly predictable in the type of plays they consider to be optimal which makes it pretty predictable to forecast the most common builds. Obviously in the lower mme fields we tend to see an obscene amount of variance and fluctuations in builds that range from random guessing to well proven models such as your own being employed. What is your opinion on how important a role game theory plays in overall approach to understanding how to build on any given day. Do you employ it at all, or do you stick with your model? Have you factored it into your model. Etc…

    A few more I am curious about. Are you willing to say which sources for projections you find to be more/most accurate or usefull. Do you take into consideration the betting markets odds on props/ teams ever, or do you consider it a distraction that has no purpose? Has there been a source or influence that has had an impact on your success? Is there a philosophical approach that you take into consideration, or is it strictly data driven choices ?

  • stv1313

    • 950

      RG Overall Ranking

    I play multiple sports and haven’t done too badly over the years. With that being said, whenever I’m asked which sport I’m best at, I usually respond with something akin to “It doesn’t matter the sport, the first month (or so) is usually the most profitable – as the inexperienced players typically feed the fish during the start of each major professional league season.”

    Have you found this to be true? Do you analyze ROI during each major sport and do you find that it declines as the season progresses? If so, do you consciously invest more during the start of a season and taper off as the season progresses?

  • factorial89

    i majored in math in college and took 2 courses in game theory.some of the stuff in this thread souns like what is called dominant strategy in nash equilibria.i have my doubts that DFS can be reduced to math and crunching numbers.

  • BrianVT

    @factorial89 said...

    i majored in math in college and took 2 courses in game theory.some of the stuff in this thread souns like what is called dominant strategy in nash equilibria.i have my doubts that DFS can be reduced to math and crunching numbers.

    You can never figure out the human element with math (the athletes), but you just try gain an edge on the competition (other DFS-ers).

  • Pidinolo

    @BrianVT said...

    You can never figure out the human element with math (the athletes), but you just try gain an edge on the competition (other DFS-ers).

    Exactly. I stated above that I found it simple to accurately predict types of builds that were most common in specific game type/ field type. Yes I could use historical data, but that lacks context. Baseball single entry is a relatively easy game to use this in. I have dozens of top tens, and several wins with negative points accrued from pitchers or 0 scores from 1 or more spots. This doesn’t always guarantee long term profitability due to unknowns within the game itself, but definitely can give you a consistent edge in single entry. It can also leaf to very polarized results vs a balanced cashing %.

    I used nash equilibrium concepts quite often and quite profitably in poker before black friday. I have really struggled to find a usefull approach capable of adapting it in dfs. I have quite often come to the conclusion that it might be limiting my possibilities vs giving me some perceived edge. It would be interesting to see the ops response…

  • CottonCombsPhD

    The 47% is happening in a NBA 20-LU game only with a cash cutline of 30%. To be clear, I have a defined setup, I don’t pay an attention to the players. I’ve collected quite a bit of data building a profile of what winning lineups look like. For instance, if you create just one optimal lineup, you see the 9 players in a configuration that generates the highest projected score. Without giving too much away, I limit the number of those players in all my lineups. So this, along with quite a few other metrics that I’ve put together are put in and then I run 20 lineups that are based on the highest score projections. Most nights I’ll have one or two players that are in every lineup. I don’t fight this, I don’t argue with the math which is based on correlations. From the sample I have for this NBA season, the daily FC projections in my sample have an R-value of .752. I don’t collect all the data, I don’t need it, my sample is big enough to know that’s very accurate. I’ve not analyzed it yet, but if the goal is to have percentage of lineups in the money, and there is a strong correlation with final scores and projections, then I would imagine there is a point of diminishing returns in terms of the number of lineups entered. Each night, there can be only so many winning lineups, so if I put my top 20 versus the top 150 (ranked in order of projection), then by definition, my top 20 over the long run should do much better than the bottom 130. Now that top 20, may look more like other lineups, but as long as they are in the money, I don’t care. I’m very interested in knowing what the breakeven is for NBA (I don’t take NFL serious, too few slates). What I mean is, there are approximately 135 days of NBA regular season this year. If you put in 150 lineups in the 3.33 game @ FD, that would be 499.5 a night (there may be a few more days, give or take, this is just an example)..so for 135 days, I have to invest $67,432.50, and get that much back to breakeven. Now I admit, it’s probably easy to get that much back, but when you do hit the big payday, now you pay taxes, then there is the cost of doing biz, you have to pay for the optimizer, then your time….maybe it’s crazy, but I would rather have an automated process that gives me 20 lineups that I can generate in 5 minutes, put those in, and probably never win, but I can cash between 40% and 50% a night, and not pay taxes, and have reduced costs for the optimizer (FC is cheaper when you can only crunch 20 LU at a time), then to me, that’s more attractive. I can fly under the radar and quietly make a few bucks each night and then over the course of a year, perhaps have a profit. So I kind of started this last MLB season. For me, the cycle will be MLB, then NBA, after 2 years of grinding away on the numbers, I’m finally figuring out the NBA. It’s funny, at the end of the day, it had very little to do with knowing all the players, the projections do that for me. I find playing the right games, with the highest cash% level is most important, and understanding how to use all the functionality of my optimizer. It took me a year to figure out how flexible my optimizer is…it’s really a powerful tool…I think if people find a stat package they like, and dump the completed player pools into a stat package, they are way ahead of everyone else…just my .2….

  • HopscotchAnon

    @Pidinolo said...

    What is your opinion on how important a role game theory plays in overall approach to understanding how to build on any given day. Do you employ it at all, or do you stick with your model? Have you factored it into your model. Etc…

    A few more I am curious about. Are you willing to say which sources for projections you find to be more/most accurate or usefull. Do you take into consideration the betting markets odds on props/ teams ever, or do you consider it a distraction that has no purpose? Has there been a source or influence that has had an impact on your success? Is there a philosophical approach that you take into consideration, or is it strictly data driven choices ?

    Yes, I consider game theory aspects. But those are automated and factored directly into the model. I don’t read Rotogrinders and see a specific player is mentioned the most and then increase/reduce my exposure. I don’t want to go into too much detail as specific elements that make a model work (or not) are a competitive advantage.

    I won’t mention specific sites or which are the best, but those that have a long history of projections and are paid for tend be higher quality. Betting odds are definitely a factor, but their importance varies between sports and what kind of bet they are on. My general philosophy is to incorporate as much data from as many sources as possible, and allow the model to determine what is statistically significant. There is some feature engineering for specific sports, as the scoring structure and some domain knowledge can be important. In terms of influence, there is a lot of similarity between DFS, quantitative finance, and algorithmic betting (especially horse racing). They can all be heavily data driven with raw money as the output.

  • HopscotchAnon

    @stv1313 said...

    Have you found this to be true? Do you analyze ROI during each major sport and do you find that it declines as the season progresses? If so, do you consciously invest more during the start of a season and taper off as the season progresses?

    This is heavily sport dependent, but ROI certainly changes as the seasons progress. In most sports, the largest changes tend to occur in playoffs, as the raw strategy of teams (and almost the sport itself) can change dramatically. Along with a smaller sample size of such games, there is a lot of opportunity to (hopefully) take advantage of those changes. I don’t invest more at any point, as the return profile remains attractive at every point.

  • HopscotchAnon

    @factorial89 said...

    i majored in math in college and took 2 courses in game theory.some of the stuff in this thread souns like what is called dominant strategy in nash equilibria.i have my doubts that DFS can be reduced to math and crunching numbers.

    Those two points seem unrelated. A dominant strategy can incorporate game theory factors through data and math. Nothing in game theory requires your views on opposing players to be based on human hunches. And I think calling anyone’s model/strategy in DFS a dominant strategy is a bit of stretch… although that is certainly the goal.

    I do agree that DFS can not be reduced to purely math and crunching numbers. There will always be statistically predictive data that is not available in a parseable format. But the trend has been, and will continue to be, toward more math and crunching numbers.

  • HopscotchAnon

    @CottonCombsPhD said...

    I would imagine there is a point of diminishing returns in terms of the number of lineups entered. Each night, there can be only so many winning lineups, so if I put my top 20 versus the top 150 (ranked in order of projection), then by definition, my top 20 over the long run should do much better than the bottom 130. Now that top 20, may look more like other lineups, but as long as they are in the money, I don’t care. I’m very interested in knowing what the breakeven is for NBA (I don’t take NFL serious, too few slates). What I mean is, there are approximately 135 days of NBA regular season this year. If you put in 150 lineups in the 3.33 game @ FD, that would be 499.5 a night (there may be a few more days, give or take, this is just an example)..so for 135 days, I have to invest $67,432.50, and get that much back to breakeven. Now I admit, it’s probably easy to get that much back, but when you do hit the big payday, now you pay taxes, then there is the cost of doing biz, you have to pay for the optimizer, then your time….maybe it’s crazy, but I would rather have an automated process that gives me 20 lineups that I can generate in 5 minutes

    Yes, there is a point of diminishing returns, almost certainly from the very beginning. Your second lineup should have a lower expected profit than your first, otherwise it should be the first lineup. But at the current entry limits, none of my lineups have negative expectation.

    If you mean more in terms of time invested by the DFS player, I certainly agree. At a reasonable scale and profit, the cost of projections / tools / Internet pale in comparison to the cost in your time, and almost any additional automation is valuable. If you can have an automated process that generates 20 lineups in 5 minutes, why not have an automated process that generates up to 150 lineups for every tournament available in 5 minutes? (I’m not sure what you mean by paying taxes – if you make any profit, it is taxable.)

  • CityBoy410

    @jjwd said...

    I’m not buying it. CEOs are sharp. Sharp guys would appreciate the chance to talk strategy in a thoughtful thread.

    Facts. Lol
    This person is just giving us their point of view and hopefully some of us non pros can use it to better our play.

  • dumbo

    • 125

      RG Overall Ranking

    • Ranked #92

      RG Tiered Ranking

    So, for the people that primarily focus on data, in nfl, do you incorporate that into stacks? Or do you spread your players out based upon projections. I can’t fathom spreading your players out on games.

    I’m usually running a 3,1 or 3,2 stack on one game or two games as my primary core and spreading and duping the other positions out on high volatile plays vs. that player’s salary. I think the only way to profit at nfl or college is to stack with a tight core. It’s very volatile but has worked.

  • jimmyquinella

    • Blogger of the Month

    Interesting information and solid questions, thanks for starting the thread.

  • TheDataDetective

    • Blogger of the Month

    Really appreciate you taking the time to do this, HA. Do you get your data directly from the authoritative sources (e.g. the sports leagues or their authorized data distributors) or from 3rd party sources?

    I wrote an MLB projections app that was successfully using data from MLB.com (i.e. their daily XML / JSON feeds) for a couple years but then they blocked access to the general public and now only distribute their data to authorized partners (teams, etc) and for a fee that is prohibitive to casual users. I suppose that someone like yourself has large enough volumes that it justifies paying the price for timely, accurate data. Curious to hear your take on this.

  • HopscotchAnon

    @dumbo said...

    So, for the people that primarily focus on data, in nfl, do you incorporate that into stacks? Or do you spread your players out based upon projections. I can’t fathom spreading your players out on games.

    If there is data that shows stacking (or explicitly not stacking) can increase the probability of a lineup winning, it should be included in the model. Raw point projections are generally just a starting point for determining exposure, but the most important one.

  • HopscotchAnon

    @TheDataDetective said...

    Really appreciate you taking the time to do this, HA. Do you get your data directly from the authoritative sources (e.g. the sports leagues or their authorized data distributors) or from 3rd party sources?

    Yes, most of the granular game data I use is directly from the leagues themselves, and what I use is all available online for free with some digging. MLB data is certainly a little frustrating, especially how every few years they change what and how they decide to release publicly, but the data is still available. I’ve found the league data very accurate, and haven’t had a need for paid game data. Although, I generally don’t play anything that requires live data, and that is when the data gets exorbitantly expensive.

  • vukrado

    What is your RMSE for your NBA model?

  • HopscotchAnon

    @vukrado said...

    What is your RMSE for your NBA model?

    I don’t have a single metric like that that would be of any utility. I appreciate the desire to compare models, but doing such a comparison is going to be pretty fruitless. Pre-season, regular season, or post-season? The past year, or past five years? Over the in-sample or out-of-sample dataset? How are players that are projected but didn’t play, or are not projected and did play, accounted for? Are known injuries included? Is the projection just raw point total expectation? The list could keep going, but unless you have the data to ensure the metric will be exactly equivalent across the different models it is not going to useful.

  • Avsrick

    Well, The one thing I learned from this thread is that my spreadsheet model would look like a 5 year old created it.

  • vukrado

    What metrics did you use after initially creating your models?

  • HopscotchAnon

    @vukrado said...

    What metrics did you use after initially creating your models?

    For the regression analysis, I use r-squared primarily, but also pseudo-r-squared and MAD based on what is being looked at. For the model as a whole in historical simulations, the final output is judged primarily on returns and Sharpe/Sortino ratios.

  • nerdytenor

    • 36

      RG Overall Ranking

    • Ranked #27

      RG Tiered Ranking

    @HopscotchAnon said...

    My cash rates (and ROI) are quite similar between the sites. I would have expected them both to be higher on DraftKings, for similar reasons as you. Namely, the multiple position eligibility and the generally increased roster flexibility allows for more unique lineups and the ability to reduce some randomness. This is also compounded by the ability to late swap with more flexibility. But I’m not sure why this is not borne out in our results.

    The added complexity of DK (which would seem to favor sharper players) may just be counterbalanced by the fact that FanDuel generally has softer pricing, so it is much easier to make a really bad lineup that uses most or all of the salary cap. (Something blenderhd has commented on in the past)

  • X Unread Thread
  • X Thread with New Replies*
  • *Jumps to your first unread reply

Subforum Index

RotoGrinders.com is the home of the daily fantasy sports community. Our content, rankings, member blogs, promotions and forum discussion all cater to the players that like to create a new fantasy team every day of the week.

If you or someone you know has a gambling problem, crisis counseling and referral services can be accessed by calling 1-800-GAMBLER (1-800-426-2537) (IL). Gambling problem? Call 1-800-GAMBLER (NJ/WV/PA/MI), 1-800-9-WITH-IT (IN), 1-800-522-4700 (CO), 1-800-BETS OFF (IA), 1-888-532-3500 (VA) or call/text TN REDLINE 1-800-889-9789 (TN).