Evidence
How our model works
Most fantasy tools ask you to take their word for it. This page is the test we ran instead, against real DraftKings contests, with the places the model is wrong shown next to the places it is right.
The test
Predicting ownership is the hard part of tournament play. It is not a forecast about football, it is a forecast about people: which players the field will pile into, and therefore which lineups will be duplicated hundreds of times and split their prize.
We hold a complete archive of real DraftKings NFL contests from the 2024 and 2025 seasons including every entered lineup, so the model can be checked against what actually happened rather than against itself.
The result
Our current ownership model was run against all 261 usable 2024 contests in the archive, covering all 18 weeks and just over 15 million entries, and scored against what the field really did.
What that last pair shows, and what it does not. It compares two averages. This time the difference is −5.2 points (95% range −6.8 to −3.5), so our top estimate now runs lower than the real top on average. On 3 October the two averages were within a point of each other. The pair also does not say the model is that close in any one contest. Contest by contest, our top estimate missed the real top ownership by 10.7 points on average (95% range 9.5 to 11.8; median 8.1). It was within 5 points in 34% of contests and within 10 points in 59%, and our most-owned player was the field's most-owned player in 52% of them.
Worse at the top of the board than on 3 October, and the model changed in between. On 3 October the same test gave a miss of 9.7 points, 38% within 5, 53% on the right player and a correlation of 0.835. Contest by contest the miss grew by 0.9 points (95% range 0.3 to 1.6) and was smaller in only 109 of the 261 contests; correlation slipped by about 0.005 (a small change, but its range stays below zero). The cause is an update on 5 October that spreads projected ownership across more teams, because real fields spread it more than our column did. It was chosen on 2024 contests, including these, and checked on 2025 contests it was not fitted on, where it lowered the error on players owned 10% or more. It also pulls down the most-owned player, which is what this test measures, so read these numbers as the cost of that trade rather than a fresh test.
The same test on 2025. Over the 272 usable 2025 contests (17 of 289 could not be scored because a team in the file had no game line that week), the correlation is 0.833. Our top estimate averages 47.8% against a real 50.7%, and it missed the real top by 9.8 points (95% range 8.9 to 10.8). Before the 5 October update the miss on the same contests was 10.7, so there the update made it 0.9 points smaller (95% range 0.2 to 1.6), the opposite direction from 2024. The two seasons disagree on this measure, and we are not choosing the better one.
FanDuel. Checking FanDuel takes more rebuilding, because the archive holds salaries but not our projections. We rebuilt player pools for 105 NFL Classic contests across 48 slates (93 from 2021 to 2025 and 12 from this season), with projections built from earlier weeks only. The rebuild does use two facts known after kickoff, which week a slate was and who did not play, which flatters it slightly. Across the players in those pools the per-contest correlation with real ownership averaged 0.69 (95% range 0.66 to 0.72, resampling slates), and our top estimate missed the real top by 11.1 points on average (95% range 8.7 to 13.5). The 5 October update lowered the average error on players owned 10% or more from 9.7 to 8.7 points; on the 58 contests it was not fitted on, from 10.0 to 9.0. This is a thin sample: 48 slates, only 15 of them from 2021 to 2023, and 61 of the 166 FanDuel contests in the archive could not be rebuilt (the export had no player ids or salaries, or the week could not be identified). Treat it as a first read.
A separate check: ownership across a contest should sum to 900%, because nine roster slots are filled by nine distinct players. Real contests average 898.5%. Our model reaches 900% by construction, so this confirms the target rather than testing the model.
Where it is wrong
On old slates it used to overstate the top of the board. The model reads pre-game Vegas totals to judge which games the field will crowd into. Those exist for recent seasons and not for archived ones, so on historical slates the term drops out and the model leans too hard on price. We found it projecting a lead back at 100% ownership, which no real contest produces. The ranking still held; the magnitudes did not. That was fixed in September, and the 10 October test above now shows the opposite lean: our top estimate runs about five points low.
Our duplication constant was wrong by more than an order of magnitude. It shipped at a value fitted before we had this archive. Measured against real duplicate counts across all 18 weeks, the true figure is roughly twenty times smaller. Every estimate that depended on it was wrong until it was corrected.
Simulated return is too noisy to rank on naively. Holding four lineups fixed and changing only the random seed, the same lineup returned anywhere from −86% to +40%. Anything that sorts lineups by the same simulated ROI it reports is sorting partly by luck. What the engine does about it: it runs thousands of draws, picks lineups using one half of them, and reports each lineup's return from the other half, so the number you see played no part in choosing it. Lineups are not ranked by top-1% finish rate.
Applies to the lineup builder (/dfs/build/) and the top plays on the optimizer (/dfs/optimizer/). Ranking method in place since 18 August 2026; checked against the production engine (build 9ee914a) on 15 September 2026. Corrected 15 September 2026: this paragraph used to say lineups are ranked by top-1% finish rate, which did not match the engine.
Raw simulated finish rates run high. We checked the simulator against 550 real DraftKings NFL Classic main contests from 2024 and 2025. In the study, raw simulated cash and top 1% rates ran roughly two to three times higher than what real contests paid out, and simulated return averaged +160% where a real-return proxy averaged −35%. The top 1%, win and return figures also did not track real results from one lineup to the next. What we changed: the lineup builder shows cash and top 10% only after a correction fitted on 2024 and checked on 2025, which it had not seen, and shows the rest as ranks that compare lineups with each other. College, FanDuel and Showdown have no measured correction yet, so they show ranks only.
Added 3 October 2026. The builder has shown corrected figures and ranks since 28 September 2026.
None of those were found by a user. They were found by testing the model against real results and publishing what came back, which is the only reason to keep an archive this size.
The field you play against
This page tests our ownership model. A tournament simulation also puts you against a field of lineups we generate, and that is a separate thing to get right. We graded it over 329 real contests against a standard written before the run. On the field the builder produces today, twelve measures are calibrated, four drifting, none clearly wrong, and two are things we do not model at all. The field scorecard is here, misses included.
See it yourself
Beat the Field is free and takes two minutes. It hands you four lineups people really entered in a real contest, hides the results, and shows you exactly where your pick finished among every entry that was actually in it. Nothing simulated.
The CFB moneyline record lists the picks our current college football model makes this season as a retrospective replay, each re-run on only the games before its kickoff and graded against DraftKings' closing odds, losses included.
Ownership figures re-measured 10 October 2026 with the current model over all 261 usable 2024 contests and all 272 usable 2025 contests, by the same method as the 15 September and 3 October runs (the original 23 August run used 257 and reported the top-of-board pair without the per-contest error). The FanDuel figures use a different method, described above, so they are not comparable with the DraftKings ones. Other figures measured 2026-08-23 and 2026-08-24 against the archive described above. Nothing here is betting advice.
Hear when we add something new
Leave your email and we will tell you when we add a new tool or more contests to play. No set schedule, no spam, unsubscribe whenever.