The Best Ball Market Is Always a Year Late
In the summer of 2022, Jonathan Taylor was the most expensive player in fantasy football. Underdog drafters took him first overall, ahead of Justin Jefferson, ahead of Travis Kelce, ahead of everyone. He finished the fantasy regular season as the 22nd best running back. The teams that spent the 1.01 on him advanced half as often as the field.
Three years later the market had seen enough. Taylor went 21st. He finished first, and the teams that had him advanced at nearly three times the rate of everyone else. By the measure everybody uses, he was the single best pick of the 2025 tournament.
He also barely made anyone any money.
Hold onto that, because it is the whole problem. I pulled every Best Ball Mania draft Underdog has released, six tournaments and 423,864 complete entries, to answer what looked like one question and turned out to be two.
Sixteen point seven percent
BBM puts you in a room of twelve. The top two rosters by total points over weeks 1 through 14 move on. Everyone else is finished in November, holding a team they will never be allowed to change.
So there is an obvious bar, and it is fixed at 16.7%. Did drafting this player push your teams above that line, or not?
That number, how far above or below 16.7% a player's teams finished, is what I will call lift. Josh Jacobs at +157% in 2022 means teams that drafted him advanced 42.9% of the time instead of 16.7%. Zero is average, negative means his drafters did worse than the room.
Advance rate is easy to measure and hard to argue with. It happens to one entry in six, it needs no model, and it reproduces almost perfectly. Nearly everyone who writes about best ball uses some version of it, and for most of this article so will I.
Here is what it says.
Your first two picks are a coin flip you are losing
Split every drafted player into five draft ranges by ADP and measure the average lift in each and they all come out at essentially zero. There is no cheap round. There is no dead zone. Across 1,569 player-seasons, no stretch of the board was systematically better than any other.
Some of that is arithmetic. Every team's eighteen picks have to add up to that team's result, so the board as a whole must average out. But now look at the median.
Draft ranges: average lift against how often a pick actually beat the bar
1,569 player-seasons, BBM I–VI pooled. Minimum 300 entries of exposure per player.
Mean lift sits at zero in every range. The median sits below zero in every range.
Every range has a negative median and a zero mean, which means the typical pick is a small loser, the average pick is break-even, and therefore the winners are enormous and rare.
It is worst exactly where the market is most confident. In the first two rounds only 38.7% of picks beat the bar, a lower hit rate than any range except the last-round dart throws.
Year by year the first two rounds went 41%, 44%, 43%, 35%, 30%, 38%. That is not a market failure. It is what the top of a draft board is. Elite players are priced for their ceiling, so most of them disappoint against that price, and the few that don't decide your season.
The first two rounds, priced against delivered
Every player with ADP inside the top 24, plotted at his price. BBM I is absent because its data dump carries no ADP column.
The vertical spread is the point: at the same price, outcomes ran from −86% to +185%.
The chart flattens what 2020 felt like. Three of the first twelve players off the board that year were Saquon Barkley, Christian McCaffrey and Michael Thomas. Barkley tore his ACL in week 2. McCaffrey played three games. Thomas played seven. Tens of thousands of tournaments were decided before October, and those drafters then watched a dead roster for eleven more weeks, because in best ball there is nothing else to do.
Six years of fighting the last war
Every year the room makes a positional bet before a snap is played: how much early draft capital goes to running backs rather than receivers. Watch what it did. 50.4%, then 46.8, 43.1, 32.0, 26.4. Four consecutive years of retreat, cutting its early running back spend nearly in half, then a partial step back in 2025.
Now watch what that was worth.
What the room bought, and what it was worth
Top: running back share of all picks in rounds 1–3. Bottom: advance lift for going zero-RB, meaning no running back inside the first five rounds.
The crowd spent four years selling running back, and the payoff for going zero-RB went from +46% to −47% while it did.
In 2022 running backs still took 43.1% of early draft capital, and going zero-RB through five rounds returned +46.3%. The crowd was too long and got punished. That was the year the top of the RB board detonated: Taylor at 1.01 finishing RB22, Najee Harris, D'Andre Swift, Javonte Williams tearing an ACL in week 4. The backs that actually won tournaments were free, and Josh Jacobs led the NFL in rushing from ADP 76.
So the market kept selling. By 2024 running backs were down to 26.4% of the first three rounds and receivers were up to 69.4%, the most lopsided board in the sample. That was the exact year zero-RB cost −31.7% and three early backs returned +61.5%.
The market spent four years correcting, and finished the correction precisely in time to be wrong in the other direction.
Early running back count, by year
Advance lift by number of RBs taken in rounds 1–5.
2022 and 2023 slope down: RB capital was a tax. 2024 and 2025 slope up: it was a discount. Nothing about how to build a roster changed. What changed was the price.
Two honest caveats. 2021 does not fit: the market was still long running back and running back was right, on the back of Taylor's RB1 season. Across all six years the correlation between RB share of early draft capital and the payoff for fading RB is +0.44, which is real and directional and nowhere near a law. And 2025 breaks the tidy mean-reversion story too, because the market did come back to running back and running back still paid. The crowd moved the right way and simply didn't move far enough.
You can watch the same instability at higher resolution across every position and price point at once.
Return on capital: mean advance lift by position and draft range
Exposure-weighted. Blue means the market underpriced that cell, red means it overpriced. Blank means too few players at that price to measure.
I checked all twenty position-by-price cells for sign consistency. The best any of them manages is four of six years. Not one held its sign five years running.
So that is the story advance rate tells, and I believed it was the whole story for about a week.
Advancing pays you twenty-five dollars
Here is the thing I had been quietly ignoring. Clearing that 16.7% bar does not win you anything. It returns your entry fee.
Best Ball Mania is four rounds, not one. Weeks 1 to 14 send the top two of twelve forward, which in the years measured here meant 672,672 entries becoming 112,112. Week 15 puts you in a group of thirteen and advances one, leaving 8,624. Week 16 does it again in groups of sixteen, leaving 539. Week 17 is those 539 playing for two million dollars. And your roster is frozen the whole way: the same eighteen players you drafted in June are the team that plays week 17.
Which means the prize pool is not distributed anything like the way advance rate is.
Where the money actually is
Official BBM prize structure. What an entry earns by the stage its run ends, and what share of total expected value sits at each stage. The first two rows are flat amounts; the last two are ladders paid by placing.
Sixty-five percent of all expected value sits with the entries that reach week 17, and one in 1,248 gets there. That last row hides its own shape. More than half the finals places pay the same $3,750, so the median finalist takes the floor.
Survive the regular season and lose in week 15, which is what happens to five of every six survivors, and you have turned $25 into $25. The entire tournament is a lottery ticket whose value sits almost entirely in a round that one entry in a thousand ever plays. And even inside that round the floor does most of the work: more than half the finals places pay the same $3,750, so the median finalist does not get rich, he min-cashes.
So advance rate is not EV. It is the price of admission to the thing that is EV. Which raises the obvious question of whether the two rank players the same way.
They do not.
The gate against the finals
Every player, plotted by how much he raised your advance rate (across) against how far he was over or under his field share among the 539 finalists (up). Both axes use complete round populations, not a sample. If the two measured the same thing, this would be a straight line.
Correlation +0.50 in 2024 and +0.38 in 2025. Of the twenty players who most raised advance rate, five made the top twenty for reaching the finals, in both years.
Take 2025. Jahmyr Gibbs raised his drafters' advance rate by +116%, one of the best marks in the tournament. Then he showed up on 15 of the 539 finalist rosters. Gibbs was drafted by 8.3% of the field, so a neutral player lands on about 45 of those teams. He landed on a third of that.
Puka Nacua did the opposite. He moved the advance needle by a comparatively ordinary +31%, and then turned up on 337 of the 539 finalists. Nearly two of every three teams playing for the title had him, against the same 8.3% field share. Kyle Pitts barely registered at the gate, at +4%, and was on 225 of them.
2024 tells the same story with different names. Jared Goff was below average for advancing, at +23%, and finished on 210 of 539 finalist rosters, tied at the very top with Chuba Hubbard. Ja'Marr Chase, who posted the best advance mark in six years at +204%, made 76.
And Jonathan Taylor, the best advance-rate pick of 2025 and the reason this article opens where it does, landed on 38 of the 539. Slightly below the share of the field that drafted him.
Why you cannot simply chase the money
At this point the obvious move is to throw out advance rate and rank everyone by dollars. That is a mistake, and it took one test to see why.
Split the entries in half at random. Compute each metric on the first half and again on the second half, then check whether the two halves agree. A measurement that cannot reproduce itself on its own data cannot rank anything. Do it forty times and average.
Can the metric reproduce itself?
Split-half reliability, 60 random splits per year, 2024 and 2025, splitting each round’s complete population. 1.0 means the two halves agree perfectly. 0 means the number is noise.
Which stage a roster reaches is highly measurable, all the way to the finals. How much it then wins is barely half measurable.
Every one of those stage measures is solid. Reaching week 15 reproduces at 0.997, week 16 at 0.983, and the finals, measured across all 539 of them, at 0.889. Who gets your team deep is not a mystery. It is one of the most stable things in this dataset, and the spread is enormous: Nacua at six and a half times his field share, Gibbs at a third of it.
What you cannot measure is the last step. Score every entry by the dollars it actually won, using real placings on the real ladder, and reliability falls to 0.448. Half of any money ranking is variance that will not repeat.
The ladder explains it. Reaching the finals gets you $3,750 and more than half the places pay exactly that, so the entire spread in prize money comes from the handful of teams that finish in the top dozen. One roster winning the thing is worth more than the other 538 combined, and nothing about a June draft predicts which roster that will be.
That is the useful line, and it is not where I expected to land. The tournament is not one lottery. It is a sequence of three measurable stages followed by one that is close to a coin toss, and almost all of the skill available to you sits in the first three.
The things that held every single year
If the money is half lottery, the sensible response is not to buy more lottery tickets. It is to fix the roster-construction mistakes that are entirely within your control, because those losses cost you at every stage at once.
So I ran the strictest test I could: take every roster-construction decision I can measure, check its sign in all six seasons, and keep only the ones that never once went the wrong way. Very little survives. That is the point.
Six for six
Every bucket that held the same sign in all six seasons. Bars are each year's advance lift.
Three of the five positives are the same idea said three ways: never let a bye week empty a slot you are required to fill.
Never start a week with an empty slot
Best ball auto-starts your best lineup. If every quarterback you drafted is on bye the same week, you start a zero in a slot the format forces you to fill. That was −22.5% on average and negative in all six years. The tight end version costs about four fifths as much and is also six for six.
The most destructive single choice in the whole dataset is not a bye at all. It is rostering one tight end: −33.1% on average, never once positive. One injury and the season ends in a way no other position can replicate.
Spread your running back byes
At the most common build, five running backs, having no two of them share a bye week was positive in all six seasons. I checked it against the obvious objection, that this is really just a proxy for owning fewer running backs. It isn't. Hold the count fixed at five and the gradient is clean: full spread beats two sharing, which beats three sharing, every year.
A third tight end, but only if the first one was cheap
Pooled across everything, three tight ends beat two in all six years. That number is a trap, and I know it is a trap because I shipped it once as a rule and had to roll it back. Split by what you paid for your first tight end and it falls apart in two of the three buckets.
Three tight ends, split by what the first one cost
Advance rate by TE count within each capital bucket, six years pooled, with the per-year sign record. Baseline is 16.7%.
Only the bottom right cell is a rule.
If your first tight end comes in round 8 or later, a third advanced at 18.2% against a 16.7% baseline, positive in every one of the six years. If you took a tight end in the first three rounds, a third is negative in four of six. Same roster shape, opposite verdict, decided entirely by price.
The interior wins, the edges kill
One quarterback or one tight end is a cliff. Four or more gives the edge back. Every profitable build lives in the middle, and running back has a ceiling too: seven or more was negative in almost every year measured.
Onesie roster counts, all six years
Advance lift by number of QBs and TEs rostered.
Stack your quarterback
Pairing your quarterback with two of his own pass catchers ran +2.1% and was positive in five of six years, missing only in 2024. That is a modest number at the gate, and it should be, because correlation is not really an advance-rate play. It is a spike-week play: a stack is how you buy the one enormous Sunday that wins a week 16 group, which is exactly the stage advance rate cannot see.
The negative case is blunter. Rosters where no quarterback had a single pass-catching teammate ran −4.9% and were positive in one year out of six. Naked quarterbacks are the mistake, more than heavy stacking is the edge.
And one popular rule that is not a rule
Three quarterbacks. The current consensus is that QB3 is correct, and 4for4's BBM VII guide makes that case directly on the grounds that four of last year's five best-performing structures ran three. Over six years it is positive in three: −0.5, −3.2, −7.5, +2.1, +3.9, +7.2. Negative in the first three years, positive in the last three, trending up the whole way. That is not a constant, it is a regime in progress being read as a law because the recent half of the sample is the half everybody remembers.
How this squares with everyone else
The public consensus holds up well. Hayden Winks' Underdog piece on five years of BBM lands on two running backs through round 6, six or seven receivers, two or three tight ends with no early bully TE, and a warning about scrolling past ADP 215. My six years of outcome data agree with all of it.
The difference is what is being measured. Most guides score structures by how many finalists had them, which is a survivorship view: it tells you what winning teams looked like, not what a choice was worth when you made it. And as the reliability test shows, finalist counts are the noisiest thing in this entire dataset. Everything here is measured at the draft, across every entry, including the 83% that go out in November.
What I can't tell you
The outcome is rebuilt, not assumed. For 2021, 2022, 2024 and 2025 I grouped the sample by draft room, ranked the twelve rosters, and flagged the top two. Checked against Underdog's own next-round entry lists, the reconstruction agrees on 99.98% of 277,908 entries. The 2023 sample is drawn per entry rather than per room, so it uses a modeled advance probability from each roster's percentile in the field; on the four years where both exist the two methods correlate at r = 0.992 to 0.995.
Samples and populations are not the same thing, and mixing them up is the easiest way to publish something false. The draft files I have are a random slice of the field, about 9% of it. The round files are the complete populations: 112,112 entries in week 15, 8,624 in week 16, 539 in the finals. So every playoff number here counts the whole round and divides by field exposure measured on the sample, which at these counts is precise to about a tenth of a percentage point. An earlier draft of this piece counted finalists inside the sample, which holds only about 50 of the 539, and that is enough to make a player look like he was on none of them when he was really on fifteen. Advance rate was checked both ways and agrees at r = 0.993.
The money figures cover 2024 and 2025 only, the years with complete round-by-round files. Placements are real, taken from each entry's actual week 15, 16 and 17 scores. Dollar levels are not: those years paid 539 finalists and the only ladder I have transcribed is BBM VII's 667-place one, so the ladder is interpolated onto 539 places and the prize totals should be read as a consistent yardstick rather than an exact payout. The structural facts quoted for those years, field size and group sizes and advance counts, are derived from the round files themselves.
BBM I is a different animal. Its dump has no ADP column, no position, no draft room id, and rounds instead of pick numbers, and its roster_points field is overwritten with a playoff-week score for teams that advanced. What it does carry is Underdog's own advance flag. Its base rate lands at 18.75% rather than 16.7%, meaning the dump does not contain every eliminated team, so treat 2020's magnitudes as the softest thing here.
Lift is raw, not adjusted. It is a comparison against the field baseline, not a regression, so a player's number carries some of his drafters' other picks with it. Treat gaps under about ten points as ties, and treat gaps in the money column with considerably more suspicion than that.
Positional finishes in the tables are weeks 1 to 14, because that is the window the advance decision is made on. They will not match full-season rankings and that is deliberate. Jaxon Smith-Njigba was the WR1 of the 2025 advance window and the WR2 of the 2025 season.
Why it stops at 2025. BBM VII drafted through the summer of 2026 and is four weeks into its regular season as I write. There is no outcome to measure yet.
The ledger, year by year
The best and worst picks of every Best Ball Mania by advance lift. Market rank is where ADP had a player at his position. The second number is where he finished through week 14.
The most efficiently priced year in the sample. Running backs took 50.4% of the first three rounds, the high-water mark the next four years would walk back from, and almost every position-by-price cell landed within ten points of zero. What moved 2020 was luck at the very top, and a rookie receiver going in round 13 who finished WR8.
The largest and cleanest dataset of the six, the complete field with Underdog's own advance flags attached. The crowd was long running back and running back delivered. The best pick in the tournament was a receiver nobody paid for: Cooper Kupp, the 19th receiver off the board, who won the receiving triple crown.
The expensive backs broke and the free ones won. It is also the one year premium onesies were correct: Kelce returned +85.6% from ADP 12.7, and taking a quarterback in the first three rounds, which meant Hurts or Mahomes, returned +42.3%. That is the best quarterback-timing reading in the sample and it never worked that well again.
The most expensive players in the draft got hurt. Justin Jefferson went 1.1 and finished WR45, second worst in six years behind only Barkley. Kupp, Chubb, Ekeler, Rodgers: the top of this board was a casualty ward and only 8 of 23 first-two-round picks beat the bar. The winners came from nowhere, Puka Nacua arriving from ADP 212. Note too what happened to the quarterback trade the crowd had just copied from 2022: taking a QB in rounds 4 to 7, the crowded middle, returned −34.3%.
The most lopsided board in the sample and the year running backs won it. Only 7 of 23 first-two-round picks cleared the bar, the lowest rate of the six. Ja'Marr Chase carried the receiver position at +185.1%, the best single advance reading in six years. And the consensus 1.01, drafted first overall in essentially every room, played four games.
The market adjusted, visibly, and it was not enough. Running backs in the first two rounds returned +37.7%, the strongest positional cell in the sample, while receivers in the same range returned −22.4%, the weakest. And Jonathan Taylor, marked down from 1.01 to pick 21 over three years, led every back through week 14, which is where this started.