← The Science Bowl Lab

Sit through an NSBA match and momentum feels undeniable. One team rattles off five, six, seven tossups in a row while the other can't buy a buzz; the room tilts, the trailing captain calls timeout, and you know a run is on. Science Bowl looks like the streakiest game in the world — and for once the eye test has real numbers behind it. But "looks streaky" and "is streaky" are different claims, and only one of them survives a shuffle.

This is the momentum chapter of the Lab, pulled out and shown in full. The question is narrow and answerable: does winning a tossup make the next one more likely, beyond the fact that the better, more-present team tends to win tossups anyway? The answer, across four leagues and twenty thousand buzzes, is no — and the way we know is worth walking through, because the naive test gets it spectacularly wrong.

01The appearance

why the eye test screams "streaky"

Start with what's real, because it's dramatic. Chop every game into its sequence of tossup winners and the runs are everywhere:

Between 70 and 78 percent of all tossups are won by a team that just won the previous one too. Runs stretch to 11 straight in NSBA and 19 straight in open play. If you defined "streaky" as "most of the action happens during runs," you'd close the case right here. The problem is that this is exactly what you'd see if there were no momentum at all — because two very ordinary forces manufacture runs on their own. First, one team is usually better (or, in NSBA, fielded the stronger lineup that night), and a better team wins more tossups, which strings together into runs mechanically. Second — subtler and famous — a naive "P(win next | just won)" is biased downward in short sequences by the Miller-Sanjurjo effect, the same selection artifact that briefly resurrected the basketball hot hand in 2018. To see momentum, you have to strip both out. A shuffle does exactly that.

02The test

hold the score, reshuffle the order

The clean null is almost embarrassingly simple. Take a real game — say a team won 20 of 34 tossups. Keep those totals fixed and reshuffle only the order of who won each one, 20,000 times. Every shuffled game has the identical final margin; all that changes is the sequence. This does two things at once, for free: it absorbs the team-strength difference (the 20–14 split is preserved in every shuffle), and because you resample the same finite sequence, it absorbs the Miller-Sanjurjo small-sample bias exactly — the bias hits the shuffled world identically to the real one. Whatever excess the real games show over the shuffled distribution is momentum and nothing else.

There is no excess.

Look at what this dismantles. In SSB, the raw "streakiness" runs about 0.50 → 0.60 — a team that just won is 60% likely to win the next, which sounds like momentum. But the shuffle reproduces 0.591 of that 0.596 from nothing but the fixed score. The entire apparent effect is the stronger team winning more often; the momentum-only residual is +0.005. NSBA goes the other way and comes out fractionally negative. Three leagues, three verdicts, one conclusion: observed ≈ chance.

03Autocorrelation says the same thing

a second lens, same verdict

If you dislike the conditional-probability framing, try the time-series one. Lag-1 autocorrelation of the winner sequence measures how much one tossup's outcome predicts the next. In SSB it comes out to a healthy-looking +0.16 — positive serial correlation, the statistical signature people point to when they say a game is streaky.

But the same trap is waiting. A pure strength difference creates serial correlation under independence: if one team wins 60% of tossups with no memory at all, consecutive outcomes are still correlated simply because the same team keeps being favored. Run the +0.16 through the order-only shuffle and almost all of it is the null. The momentum-only excess is +0.02 — not significant, indistinguishable from zero. Both lenses, the conditional probability and the autocorrelation, agree once you subtract the part that independence already explains.

04Blowouts are one long run — mechanically

the strongest-looking streak is a consequence, not a cause

The best-looking evidence for momentum is the blowout: a team catches fire, reels off fifteen straight, and wins going away. And the correlation is genuinely strong — a game's margin and its longest run travel together at +0.67 to +0.78 across leagues. Surely that is momentum.

It's arithmetic. If you win 25 of 30 tossups, you are forced to string long runs together — there simply aren't enough losses to break them up. The longest run in a lopsided game is bounded below by the margin; a 25–5 game cannot help but contain a big streak, whether or not any buzz "caused" the next. The run doesn't produce the blowout; the blowout produces the run. This is why the order-only shuffle is the right tool: it holds the 25–5 fixed and asks whether the real ordering clusters more than random orderings of that same 25–5. It doesn't. The margin–run correlation is real and the causation points backward.

05The one real effect: cold packets, not cold teams

something does cluster — it just isn't momentum

The shuffle isn't a blunt instrument that finds nothing everywhere. There is one place where NSBA departs from its null: dead tossups cluster — the questions nobody answers arrive in bunches (p = 0.017). Stretches where the buzzers go quiet on both sides are more common than random ordering predicts.

But read what that is. A dead tossup isn't a team going cold; it's a question being hard — and hard questions travel in packs because packet difficulty is lumpy. A run of obscure Earth & Space or a dense computational set silences both teams at once. That's a property of the material, not the players' momentum, and it's the opposite of a hot hand: it's neither team scoring, together, because the questions got hard. Tellingly, SSB shows the reverse — its dead tossups are more spread out than chance, consistent with escalating-difficulty tournament packet design smearing the hard questions across rounds rather than bunching them. Two leagues, opposite clustering, both explained by how packets are built — and neither by anyone catching fire.

06What it means for who wins

streakiness adds no variance to the outcome

Here is the part that matters for anyone modeling games or building a team. Because the shuffle preserves each game's final margin exactly, and because the real ordering is statistically indistinguishable from a random ordering of that margin, streakiness adds no variance to who wins or by how much. Momentum, if it existed, would be a source of outcome uncertainty a model would have to account for — a reason a close game could snowball. It isn't one. Rearranging when the runs happen never changes the final score, and the runs aren't arranged in any special way.

So the tossup sequence behaves like a series of independent, talent-weighted coin flips: each buzz is decided by who's better and who's present, with no memory of the last one. You can throw the whole game's outcome distribution off the per-tossup win probabilities and the fixed margin, and lose nothing by ignoring streaks — which is exactly what the rest of the Lab does.

None of this says the feeling is fake — the room really does tilt, and a seven-tossup run really does happen. It says the run carries no predictive information the score doesn't already contain. The next buzz doesn't know the last one happened.

07Methods & honest caveats