The Season in Data

The Parity Principle

The lesson: Science bowl is a high-variance game where the skill gap shows up as margin, not as predictability — so a single-elim bracket is a coin-flip carnival, and earning a top seed barely helps.

Watch enough NSBA and two facts fight each other. The scores are brutal — most games are not close, they are routs. Yet the winners are shockingly hard to call in advance. Both things are true at once, and reconciling them is the whole theory of this sport. We call it the parity principle: enormous variance compresses everyone toward the middle when it matters most.

Start with how a typical game looks. The average final margin this season was 45.4 points, with a median of 29 — a gap between mean and median that screams a long right tail of demolitions. Nearly half of all games, 45.2%, were decided by 40 or more. Only 14.3% finished within 10 points. In most sports a one-score game is the modal outcome; in science bowl the modal outcome is a beatdown. The buzzer race is winner-take-all on every tossup, so small edges in speed and accuracy snowball into scoreboard avalanches.

Naively, lopsided scores should make prediction easy — if teams win by 45, surely you know who is better. Here is the twist that defines the sport: they don't make it easy at all. When we regress pre-game team ratings on eventual margin, those ratings explain only about 6% of the variance in the final score (R² ≈ 0.06). That is the number that runs this league. Note the distinction from a descriptive stat like "gets explain the bulk of margin" — of course they do, that's tautological, gets are the score. The predictive fact is that what we knew going in barely moves the needle. Roughly 94% of who-wins-by-how-much is noise relative to our best forecast: which questions came up, who slept well, who caught a hot buzzer streak.

Put those two findings together and the paradox resolves. Blowouts are common, but which direction a given blowout points is close to a coin flip between comparable teams. The skill gap is real — it just expresses itself as the size of wins over a long season, not as reliable game-to-game outcomes. Averaged over a schedule the strong separate from the weak; on any single night, a 40-point swing is one bad packet away.

What that does to a bracket

Feed that 6%-signal reality into a 50,000-run Monte Carlo of the eight-team playoff and the carnival becomes literal. The title is a near-uniform lottery. The No. 1 overall projection, James W, wins it all just 15.9% of the time. The bottom playoff team, (Stitch)², still cashes 10.1% — barely a coin flip's worth behind. The entire eight-team field is bunched inside a razor-thin 10–16% band. In a league with real parity you'd expect the top seed to lap the field; here the best team is only about 1.6× more likely to win than the worst qualifier.

Look closely and the seeding almost inverts. The East's No. 1, dan.k.memes, projects below its own conference's No. 2 and No. 4 at 11.5%. That isn't a bug — it's variance eating the value of a good record. When one game is a coin flip, the reward for stacking a great regular season is a marginally better coin, not a bye to the trophy.

Volume Is King

The lesson: Games are won by getting to the buzzer first and often — out-buzzing your opponent moves the scoreboard roughly six times harder than out-disciplining them.

Every science-bowl team tells itself a story about how it wins. Some preach discipline: don't neg, don't gamble, let the other team hang itself. Some preach the bonus: convert every toss-up into a clean four-for-four. Both stories are real, and both are second-order. When you regress a team's point margin across all 42 NSBA4 games onto just two standardized inputs — its toss-up gets-differential and its neg-differential — one coefficient towers over the other.

Standardized regression coefficients on point margin, all 42 games. Bars scaled to the larger absolute effect. Model R² = 0.885.

Read those numbers plainly. A one-standard-deviation edge in toss-ups is worth +0.875 standard deviations of margin; the same-sized edge in avoiding negs is worth only −0.152. That is a ~6× gap. Buzzing more is not a little more important than staying clean — it is the whole game, and neg-avoidance is a rounding correction on top of it. Together the two inputs explain 88.5% of margin variance, which means almost nothing about who wins a science-bowl match happens away from the toss-up battle.

The standings agree with the regression. Sort all fourteen teams by raw toss-up volume and the leaderboard reorganizes itself into the playoff picture. The two heaviest-buzzing rosters, James W (75 gets) and mingle/Yunyi (73), sit at 5–1 and 4–2 with the conference's fattest point differentials; cryo (70) is right behind at 4–2. Drop to the bottom of the volume table — ChessFun (51), GidTheKid2 (53) — and you find 2–4 records and point differentials underwater by 170-plus. Across the league, gets correlate with wins at r = 0.57 and with point differential at a near-mechanical r = 0.93. Negs, by contrast, correlate with wins at only r = −0.31.

Teams sorted by toss-up gets (top 10 of 14). Volume tracks point differential almost perfectly; wins wobble around it because schedule and close-game luck get a vote.

Two exceptions prove the rule rather than break it. Connor Chang buzzes like a contender (67 gets, third-fewest negs) yet sits 2–4 — a team that piled up volume in the games it already led and got unlucky in the ones it didn't; its +84 differential says the record understates it. And dan.k.memes went 5–1 on only 60 gets and a differential of −6, the purest close-game record in the league — winning the tight ones is real, but it is variance you cannot bank on, not a repeatable edge. Meanwhile jonathan is the cautionary tale the neg-avoidance crowd forgets: 36 negs, dead last, 1–5. Discipline matters most precisely when it collapses.

The Neg Tax & the Patience Premium

The lesson: A tossup pays +4 and costs −4, so an early buzz needs ~50% accuracy just to break even — don't buy the power unless you're sure.

Every buzz in NSBA4 is a bet with fixed, symmetric stakes: sit on a correct answer and you bank +4 points; jump the gun and neg, and you hand your opponent −4 plus the rebound. That symmetry is the whole game. Because the reward and the punishment are equal and opposite, the expected value of a buzz collapses to a single lever: your probability of being right. Buzz with accuracy p, and your expected haul is 4p − 4(1−p) = 8p − 4. That line crosses zero at exactly p = 0.50. Below coin-flip confidence, you are literally paying the house to answer.

We call that price the neg tax, and the cleanest way to see it is to sort every buzz by when it happened. Our timing variable is buzz_pos — the fraction of the question that had been read aloud at the moment the buzzer went down. A value of 0.10 means you interrupted after the first sentence; 0.95 means you waited for the giveaway. Bucketing all 1,146 buzzes by buzz_pos and averaging their point value produces a strikingly monotone curve:

Expected points per buzz, by how much of the question was read

The message is blunt. Buzzing anywhere in the first half of the question is a near coin-flip that barely clears break-even: the 0.25–0.50 bucket negs 48.8% of the time (accuracy just 51.2%) and returns a rounding-error +0.09 points — nine hundredths of a point for the privilege of showboating. Wait until the last quarter and the neg rate collapses to 13.9% while the payout climbs to +1.96, more than twenty times the early-buzz yield. That gap — roughly +1.9 points per tossup between the patient buzz and the speculative one — is the patience premium, and over a season it is the difference between a contender and a cellar-dweller.

Which brings us to the players who paid the tax. The leaderboard of the neg-prone reads like a cautionary tale — ten players who negged on more than a third of their buzzes, several of them landing in the red for the season:

Note the split personalities here. Abhinav A's four gets against seven negs is a −12 hole dug almost entirely by speculative buzzes with a mean buzz_pos of 0.68. But potential man and pancreas uvula appear on this list too — and finished +40 and +32. The distinction is volume with hit-rate: they negged a lot in absolute terms because they buzzed a lot, but converted well over half their attempts. Aggression isn't the sin; aggression without the underlying accuracy to clear the 50% break-even line is. League-wide the numbers land on the right side of the ledger — a 63.9% get rate against a 19.6% neg rate — which is exactly why the average buzz is worth a healthy +1.72 points.

So the strategic prescription writes itself, and it is the least glamorous advice in quizbowl. The buzzer does not reward bravado; it rewards certainty, and certainty almost always arrives later in the question. Unless you can honestly clear coin-flip confidence — and remember, the pros who buzz early are only doing so when they're already sure — let the clue keep reading. The power is a luxury good. The neg tax is the bill.

Every Subject Has a Personality

The lesson: Buzz recall subjects (Earth/Space, Biology) early and calculation subjects (Math) late — and value a reliable Math or CS answerer, not because those questions are scarce (every subject supplies equally) but because they go dead far more often.

Ask a room of science-bowl players "when should I buzz?" and you'll get one answer: "as early as I'm sure." That's true but useless, because the moment you're sure arrives at wildly different points depending on what kind of question is being read. A tossup is not a neutral container. Its subject dictates its architecture, and its architecture dictates its tempo. Across NSBA4's 1,331 buzzes, the fraction of the question read at the moment of the buzz — buzz_pos, where lower means earlier — splits the six subjects into two distinct temperaments.

Recall subjects get buzzed early. Earth/Space (mean buzz_pos 0.779) and Biology (0.787) are the twitchiest categories in the game: a single diagnostic clue — a mineral's cleavage, an organelle's function — is enough to commit, because the answer is a fact you either hold or you don't. There's nothing to derive, so waiting buys you almost no additional certainty. Calculation subjects get buzzed late. Math sits alone at the back at 0.930 — players hear nearly the entire question before committing, for the obvious reason that you can't buzz on a math tossup until you've actually done the math, and the problem isn't fully specified until the last clause. Physics (0.871), Chemistry (0.856), and Computer Science (0.850) live in the sensible middle, blending a recallable front half ("this scientist," "this reaction type") with a computational or multi-step back half.

Mean fraction of question read at buzz, earliest → latest. Bars scaled to the slowest subject (Math = 100%).

Specialists vs Generalists

The lesson: It's not the shape of your scoring that wins games — it's the size. Own three categories or cover all six; either works, as long as the total is there.

Every roster-construction debate in NSBA eventually collapses into one question: do you draft a spike — two or three elite buzzers who bury a handful of categories — or do you draft coverage, a balanced lineup that answers a little of everything? The intuition cuts both ways. Specialists let you dominate the subjects you own; generalists insure you against the night your ace goes cold or the packet skews away from your strength. Both stories sound right. So we measured which one actually shows up in the standings.

The natural yardstick is the Herfindahl-Hirschman Index (HHI) — borrowed from economics, where it measures market concentration. For each team we take its distribution of gets across the six science-bowl subjects, square each subject's share, and sum. A team that splits its gets perfectly evenly across all six lands at the theoretical floor of 0.167 (a pure generalist); a team that funnels everything through one or two subjects climbs toward 1.0 (a pure specialist). Higher HHI = more lopsided.

Compare that to the stat that does move the needle. Raw firepower — total gets — correlates with wins at r = 0.569, accounting for about 32% of win variance (r² ≈ 0.32). That is a 160-fold gap in explanatory power between how much a team scores and how evenly it scores. Firepower is the engine; concentration is the paint job. And crucially, HHI and gets are themselves uncorrelated (r = −0.08), so this isn't a case of one hiding the other — balance and volume are simply orthogonal, and only volume is wired to the scoreboard.

The full league, sorted by concentration

Read the extremes and the point makes itself. The most concentrated team, GidTheKid2 (HHI 0.200, a Physics spike), went 2 wins — but not because it was lopsided; it simply had the lowest firepower in the field at 53 gets. The most balanced pair sit tied at 0.168: cryo and jonathan. cryo turned its even spread into 4 wins on 70 gets; jonathan managed just 1 win on 55. Same shape, opposite records — because the gap between them is volume, not balance. And the two best teams in the league, James W and dan.k.memes (5 wins each), sit on opposite halves of the HHI table. One got there through a Chemistry-heavy build with league-high firepower; the other through an even Earth/Space-led spread. Both routes led to the same place.

Notice too how compressed the HHI column is — the entire league lives between 0.168 and 0.200, hugging the 0.167 balance floor. Even NSBA's "specialists" are, by any absolute standard, fairly diversified; nobody is running a true single-subject monopoly. That narrow spread is itself the tell: teams already self-correct toward coverage, and the small residual differences in shape that remain carry no signal about winning.

One honest caveat: with only 14 teams and a season this short, a near-zero correlation is measured with wide error bars — we cannot rule out a small real effect for balance in some matchups, e.g. against an opponent whose strengths perfectly mirror your one weak subject. But the burden of proof runs the other way. The data offers no evidence that chasing category coverage buys wins, and strong evidence that accumulating raw scoring does. When in doubt, draft the points.

Home Court Is Real

The lesson: Hosting is a real, sizeable edge in this format — and if you get to ban a category as the home team, ban the one your opponent is best at, not the one you fear most.

"Home court" sounds like a joke in a game played over the internet — there is no crowd, no travel, no hostile gym. But NSBA4 hands the host team a concrete structural advantage: the home side bans one subject category before the match. That is not ambience; that is a rules lever. And the twelve games that carried a home designation this season point the same direction, hard: home teams went 9–3 and won by an average of +31.2 points.

Why would a category ban be worth this much? Start with the mechanics of a science-bowl match. Points are not spread evenly across a roster — they concentrate in specialists. A team's Biology ace or Math closer can single-handedly account for a huge share of its gets. When the host bans that subject, they are not shaving a few percent off the opponent's expected output; they are deleting the opponent's single best scoring channel and forcing their points to come from thinner, contested categories. It is the science-bowl equivalent of taking away a team's first option and daring the bench to beat you. Comfort and scheduling plausibly add a little on top — hosts pick slots that suit their lineup and play the version of the game they prepared for — but the ban is the mechanism doing most of the lifting.

The margin distribution tells the story better than the average does. Of the twelve host games, nine were home wins and several were routs: home teams posted margins of +46, +48, +54, +80, and a comically lopsided +118. The three home losses were comparatively tame (−12, −14, −24). When the ban works, it doesn't just tilt the game — it caves it in. The median home margin was a still-emphatic +26, so this isn't one blowout dragging the mean around.

Even discounted for sample size, the effect is large enough to matter — and it did, in a place that cost real accuracy: the betting model. Our pre-game line model treated matchups as essentially venue-neutral, pricing games off team strength and projected lineups. That is a defensible default in most sports, where home edge is a few points. Here it was a systematic leak. The model repeatedly hung road favorites that the format then upset, because a nominally stronger team walking into a hostile category ban is not, in fact, the favorite it looks like on paper. A structural +31-point host margin is exactly the signal that predicts those upsets — and a line that ignores it will keep handing them out.

So the correction runs both ways. As a modeler: add a host term. Even a conservative shrink of the +31 estimate toward, say, a +10 to +15 prior would have flipped several mispriced road favorites and stopped the bleeding. As a captain: hosting is not a formality, it is a weapon. When you get the ban, resist the instinct to ban the category that scares you. Ban the category your opponent is best at — strip their specialist, force their points into contested air, and let the +31 do the rest.

The Draft Signal

The lesson: The combine is the single best thing you can draft on — but it's only about 40% of the story, so draft on ability and then win the season on usage and availability.

Before a single tossup was read, every player carried a number: theta_overall, an IRT (item-response-theory) estimate of raw ability distilled from the pre-season combine. It's the closest thing NSBA4 has to a scouting grade — a single latent skill score fit from how each player performed against calibrated question difficulty. The question this chapter answers is the only one that matters on draft night: does the combine actually predict who scores?

The honest answer is yes, but noisily. Across our 20-player evaluation sample, the correlation between combine theta and total season points was r = 0.40. That's a real, positive, useful signal — ability is not noise. But squaring it tells the sobering half of the truth: theta explains only about 16% of the variance in season scoring (R² = 0.40² ≈ 0.16). Roughly five-sixths of what separates a 96-point season from a 0-point season lives outside the combine grade.

What the other 84% is

The missing variance isn't random mystery — most of it has a name: usage. Count each player's gets (tossups they actually got to answer) and regress points on that instead of ability, and the correlation rockets to r = 0.92 — usage alone explains 84% of scoring. Points are, to a first approximation, opportunities times conversion, and the combine measures conversion while saying almost nothing about opportunity. Tellingly, theta predicts a player's own gets at only r = 0.38: being good does not guarantee you the buzzer. That gap is where role, team context, and attendance live — a strong player buried behind a faster teammate, or one who simply didn't dress every week, scores like a scrub no matter what the combine said.

Who beat their grade — and who didn't

Fit the season's line (points ≈ 32 + 24·theta) and the residuals expose the draft's biggest hits and busts. The over-performers share a trait the combine can't see: volume. Lucas W entered as a middling +0.38 ability and led the sample with 96 points — because he saw 27 gets, more than anyone. Rohan G, same +0.38 grade, banked 76 on 26 gets. Riyan N graded below replacement at −0.71 yet posted 44, converting a modest 11 chances at a blistering 4.0 pts/get. The busts invert the pattern: Sanjay O carried a respectable +0.40 grade into a 0-point season on just 7 gets, and Abhinav A and Sam E went negative — theirs is a negs problem, ability that never turned into net points.

Notice what the over- and under-performers have in common with their own grade: almost nothing. Every top over-performer got there on gets, not theta; every bust starved on gets regardless of theta. Aldric B is the exception that proves the ability rule — a genuine +0.66 talent who converted his 9 chances at a sample-best 4.0 pts/get but simply never got the volume to cash it. His per-buzz efficiency was the combine paying off; his point total was the schedule and the lineup card refusing to let it.

So the combine earns its place as the top draft input — a 0.40 correlation is more predictive signal than any other single number you'll have on the clock, and ignoring it means drafting blind. But treat it as a prior, not a prophecy. Draft the ability, then coach the two levers the combine is deaf to: make sure your pick gets the buzzer (usage/role) and make sure they show up (attendance). Do those and a +0.38 grade wins you a scoring title. Skip them and a +0.40 grade scores zero.

Who Produced — The Leaderboards

The lesson: "Best player" is not one question but three — most points, most points per buzz, and best-in-subject — and the three lists barely overlap, because volume, efficiency, and leverage are different skills that fit different roster roles.

Every fantasy argument in this league eventually collapses into a definition fight. Who was the best NSBA4 player? The honest answer is a question of its own: best at what? Points are the currency, but points come from two multiplied ingredients — how often you buzz and how much each buzz is worth — and a subject king wins by owning a scarce category rather than by piling up raw totals. Change the lens and the names on top change with it. That is not a flaw in the data; it is the whole scouting problem.

Lens 1 — Volume: the points kings

Total points is the box-score line everyone reads first, and it rewards usage: buzz a lot, on a good team, and the counting stat climbs. Kian D ran away with it at 136 points on a league-high 52 buzzes — but note he was also efficient (2.62 points per buzz), which is what separates a volume star from a chucker. Behind him, akhil (112) and vishnu m (104) show the archetype's tax: akhil's 112 points came with 9 negs (an 18% neg rate), the price of a trigger finger that also buys you a category crown.

Lens 2 — Efficiency: points per buzz

Now divide points by buzzes and a different roster surfaces. Points-per-buzz (PPB) measures conversion — how much value you extract from every trigger you pull — and it is the metric that separates a sniper from a volume shooter. The leaderboard flips: uddip k tops it at 3.33 PPB with a spotless 0% neg rate, and snapninja (3.00) and Advik S (2.80) join him — none of whom cracked the top-five in raw points. These are low-usage, high-purity buzzers: they wait for a lock and cash it. The catch is sample size — uddip's rate rides on just 12 buzzes, so we floor the list at a minimum buzz count; a tiny denominator can manufacture a gaudy ratio the way a 2-for-2 night manufactures a 1.000 batting average.

The tell is in that last row. Kian D is the only name that appears near the top of both lists — 2.62 PPB is elite conversion, and he sustained it across 52 buzzes, more than four times uddip's volume. Efficiency at low usage is easy; efficiency at high usage is what an MVP looks like. Meanwhile vishnu m (1.96) and Edward (1.87) prove the converse: you can finish top-three in points while converting below the league's 2.0-ish bar, because sheer volume papers over a thinner margin per buzz.

Lens 3 — Leverage: the category kings

The third lens ignores totals entirely and asks who owned a subject. With six categories and roughly even supply, controlling one is a strategic asset — it lets a captain punt a weakness elsewhere. Note that these crowns spread the wealth: five different players hold the six kingships, and only akhil doubles as a top-scorer. snapninja, a Lens-2 efficiency darling barely visible in the points table, is the most-in-demand Earth/Space buzzer in the league (14 gets). Leverage is where the specialists get paid.

The Field — A Team-by-Team Season

The lesson: A seeding tiebreak is an incentive in disguise. When ties are broken by total points scored — not point differential — the rational play is to keep buzzing whether you are up 100 or down 100, because coasting can cost you a playoff berth you already thought you had earned.

Fourteen teams, six games each, two conferences, eight playoff tickets. On paper it is a simple bracket. In practice, NSBA4's seeding rule quietly rewrote how the back half of the season was played. Berths are ordered first by wins, and ties are broken by cumulative points scored across the season — not by point differential, not by head-to-head. That one design choice is why the standings below do not sort cleanly by how "good" a team looks, and why two teams with identical 3–3 records could finish on opposite sides of the cut line separated by nothing but offense.

Start with the paradox at the bottom of the East. Fez_Keyreb scored 554 points — more than three of the four East teams that made the playoffs (sumin 504, xpoes 488, and would-be fifth anurag 468) — and still went home at 1–5. Points are a tiebreak, not a ticket: wins gate the field first. Fez piled up a huge scoring total across six competitive losses (point-diff just −22) and none of it mattered, because the rule only lets total points settle ties within a win tier. The mirror image sits in the West, where (Stitch)² rode a −160 differential and a league-ugly 29 negs into the four-seed on the strength of a 3–3 record alone. Same rulebook, opposite outcomes.

Eastern Conference

Western Conference

Read the tiebreak in motion. In the West's 4–2 tier, cryo and czz arrived with identical records, and the two-seed went to cryo purely because it scored 562 to czz's 472 — a 90-point gap with no bearing on defense, schedule, or head-to-head. In the East's three-way 3–3 pileup, the entire playoff picture hinged on 36 points of offense: sumin (504) and xpoes (488) advanced; anurag (468) did not. Anurag's differential (−22) was not meaningfully worse than xpoes's (−2), but the rule does not read differential. It reads the scoreboard's left-hand column, and anurag left too many points on the table.

One honest caveat before the bracket. It is tempting to call the tiebreak "unfair" to a Connor Chang (+84 diff, home before the cut) or a Fez_Keyreb (554 points, one win). But recall the descriptive engine underneath all of this: gets alone explain 88% of a game's final margin (standardized β = 0.88), so total points scored is really a proxy for total questions answered — a clean, additive measure of production. The rule is not arbitrary; it just optimizes for a different thing than differential does. Whether that thing is the right thing is the argument the standings above will keep starting.

The Championship Simulation

The lesson: In a high-variance game, a single-elimination bracket is the great equalizer — the best team is favored but nowhere near safe, so the smart money is always "field over chalk."

To close the season we ran the playoff bracket 50,000 times. Not a napkin projection — a full Monte-Carlo, wiring the eight qualifiers into the exact same tossup-contest line engine that produced every game line this year. Each simulated game samples a margin, resolves a winner, and advances the bracket: quarterfinals inside each conference, then the two conference finals, then a Grand Final between the East and West champions for the $1,000 prize. Fifty thousand times. What comes back is less a prediction than a stress test of how much the regular season actually tells us — and the answer is humbling.

The bracket

The odds are flat — strikingly flat

Here is the whole tournament in one chart: every team's probability of walking away with the $1,000, out of 50,000 runs.

Title probability across 50,000 simulated brackets. Bars scaled to the favorite (James W = 100%). Note that a perfectly random field would sit at 12.5% each — everyone is within striking distance of that line.

Look at the ceiling and the floor. The favorite, James W, tops out at just 15.9%. The longest shot, (Stitch)², still clears 10.1%. The entire eight-team field is bunched inside a 5.9-point band, and six of the eight sit between 11% and 14%. For context, if the bracket were decided by a coin flip at every round, each team would own exactly 12.5%. We are barely off of that. The engine, handed a full season of ratings, can separate the best team from the worst by fewer than six percentage points of championship equity.

Watch the seeds scramble: the E1 top seed (dan.k.memes, 11.5%) is actually the sixth-likeliest champion, sitting behind the E4 four-seed xpoes. Seeding barely survives contact with the variance.

Why the great equalizer wins

The flatness isn't a bug in the simulation — it's the deepest truth in this dataset made visible. Recall the two faces of the margin. Descriptively, the outcome of a game is almost fully explained by what happens in it: gets and negs account for 88.5% of the variance in point margin (standardized β of 0.88 on gets). But that's tautological — of course the team that buzzes more wins. The number that matters for prediction is different and brutal: pre-game ratings explain only about 6% of margin variance. Everything else — who happens to show up, which questions get read, who gets hot — is noise we cannot forecast from the standings.

The Ten-Second Version

If you remember nothing else: out-buzz the other team (volume beats everything), don't buy the power unless you're sure (the neg tax is brutal), and accept that in a game this high-variance, the bracket is a coin-flip carnival where the top seed is favored but never safe.

Methodology: buzz-level data scraped from every scoresheet; margins decomposed with standardized OLS; ability from an IRT model on the preseason combine; win probabilities and the championship bracket from a subject-resolved tossup-contest engine (the same one behind the weekly lines), Monte-Carloed 50,000 times. Descriptive vs predictive claims are flagged throughout. Some splits (home court) are small-sample and marked as suggestive.