Grand sumo, every top-division final day, 1958 to 2026
The Eighth Win
Sumo's 2011 match-fixing purge left no mark on the final day. A wrestler who needed his eighth win beat an opponent who did not 62% of the time in the decade before it and 61% in the eight years after, against about 49% expected both times.
Fifteen bouts, and eight wins is the line: it means promotion, or at least no demotion, and near the foot of the second division it can be the difference between a salary and none. In 2002 two economists showed that wrestlers on seven wins going into the last day won far more often than they should, and that the favour seemed to be returned the next time the two met. In 2011 police found text messages arranging bouts, a tournament was cancelled, and 23 wrestlers were expelled. This page asks what the final day looked like afterwards. Every figure is recalculated in your browser from 197,779 bouts in 410 tournaments.
Unpacking the data…
Sixty-eight years of final days
Top: on the last day, how much more often the wrestler on seven wins beat an opponent who was not, than a rating model built from every other bout expected. Bottom, from a different site and a different unit: how many more wrestlers finished 8-7 than 7-8. With no line at eight, both would sit on zero. Whiskers are 95% intervals.
Hover a period.
1989 to 2000
The years Duggan and Levitt studied. The peak on both measures.
2000
A former wrestler, Keisuke Itai, says publicly and in a book that most bouts are fixed. Out of the paper's sample, and the sharpest fall in the record follows it. Then it creeps back.
2011
Text messages turn up in a police inquiry into gambling. The March tournament is cancelled, the first since 1946, and 23 wrestlers are expelled. Neither measure moves.
2019
The final-day schedule starts pairing most 7-7 wrestlers with each other, which removes most of the bouts this page measures. See below.
Before and after the purge
The decade before the scandal against the eight years after it, stopping where the schedule changed. Two measures, from two sites, kept apart and never averaged. If the purge had cleaned up the last day, both should drop.
Measure A · bouts, sumo-api.com
Wrestler on 7 wins, last day
Win rate against an opponent not on 7 wins, and what the rating model expected.
computing…
Measure B · final records, SumoDB
Finished 8-7 against 7-8
Share of wrestlers with a full fifteen bouts who finished on each record.
computing…
The controls
Each of these could catch the method lying. Their verdicts are computed live; if any fails, the headline at the top is withdrawn and says so.
Control 1 · reproduce the paper first
Duggan and Levitt's own years, rerun
January 1989 to January 2000, their published figures against this pipeline. Their day-by-day figures are the raw excess over 50%, column 1 of their Table 1, so the same is computed here.
computing…
Control 2 · same day, nothing at stake
Wrestlers on 6-8 and 8-6
Judged by the same rating model on the same last day, against the same kind of opponent. They have no eighth win to chase, so their excess should be near zero. If the model flattered anyone in the middle of the table, it would show here.
computing…
Control 3 · a second expectation
Rank instead of ratings
The 7-7 excess again, with the expected rate taken only from where each man stood on the banzuke, fitted per period. It shares nothing with the Elo model but the outcomes.
computing…
Control 4 · is the model calibrated?
Every other late bout
Days 11 to 15, nobody on the margin. Expected against actual for the favourite, period by period. A model that is off here cannot be trusted on the 7-7 bouts.
computing…
The rating model's memory
Change how fast the ratings move
Elo's K sets how far one bout moves a rating. Everything that depends on the model, including the headline and controls 2 to 4, is recomputed as you drag.
10
K = 10 was chosen before any control existed, as the value that best predicted ordinary bouts. When this page was built, the headline stood for every K from 4 to 16 and withdrew itself from 18 up, where the ratings swing so far after each bout that control 4 fails, and control 3 with it a little higher. The purge comparison itself moves by less than a point across the whole range.
The schedule changed the experiment
Who fights whom on the last day is set by the judges the evening before. Until 2018, most wrestlers on 7-7 met someone who was already safe or already beaten. From 2019 most meet another 7-7, where both need the win and neither can afford to give it away.
Hover a year.
That is a measurement problem as well as a policy. It halves the number of bouts measure A can use, and a 7-7 against 7-7 bout sends exactly one man to 8-7 and one to 7-8, which pulls measure B towards zero whatever anyone does. So the comparison above stops at November 2018, and the years since are shown on their own:
Bought, or just wanted more?
A wrestler with more at stake may simply try harder. Duggan and Levitt's answer was the rematch: if the eighth win had been bought, the man who sold it should get it back the next time the two met. They found exactly that. It is rerun here, with two comparisons the paper did not have.
computing…
Hunting the instrument
The headline rests on knowing, for every bout since 1958, who fought, who won, and what each man's record was that morning. These are the places that could have gone wrong, and what each one did.
Methods
Sources
sumo-api.com, the banzuke endpoint (/api/basho/<YYYYMM>/banzuke/<Makuuchi|Juryo>), read on 2026-10-14: every Makuuchi and Juryo wrestler in every tournament from March 1958 to September 2026, with each day's result and opponent. 826 requests. No licence is stated; the site asks only for responsible use. Results are factual records, and this page ships only results, ranks and ring names.
SumoDB (sumodb.sumogames.de), the final-day results page of each tournament, read on 2026-10-14: the "Yusho Arasoi" block lists every top-two-division wrestler with his fifteen-day string of wins, losses, defaults and absences. 413 pages. No licence is stated. Used only for measure B and the cross-check.
Japan Sumo Association (sumo.or.jp), the hoshitori chart behind its results page, for the September 2026 tournament only, used as an official spot check.
Duggan, M. and Levitt, S. D. (2002), "Winning Isn't Everything: Corruption in Sumo Wrestling", American Economic Review 92(5), for the published figures control 1 reproduces.
What was measured, and how
On the margin means still able to reach eight wins and not there yet: on day 15, exactly seven wins; on day 14, six or seven; and so on, as in the paper. Absences count against a wrestler, as they do on the banzuke. A bout counts for measure A only when exactly one of the two is on the margin.
Expected comes from an Elo rating run through every bout in order from 1958, starting everyone at 1500 with K = 10 unless you move the slider. The bouts under test (late bouts with exactly one man on the margin) are predicted but never learned from, so a thrown bout cannot inflate the rating that judges the next one. Bouts decided by default are excluded from everything.
Measure A is actual minus expected wins, as a share of bouts. Its interval is the binomial spread implied by the expected probabilities.
Measure B is the share of wrestlers who fought all fifteen days (defaults included, absences not) and finished 8-7, minus the share who finished 7-8. It uses no model at all.
Periods: the purge comparison is March 2000 to January 2011 against May 2011 to November 2018. March 2000 starts the decade after the paper's sample; May 2011 is the first tournament after the cancellation; November 2018 is the last before the schedule changed.
The headline verdict calls a change "no measurable change" only when both measures' 95% intervals for the difference include zero.
What this does not say
Not that any bout was fixed. An excess is a pattern across hundreds of bouts. It cannot point at one, and nobody named on this page is accused of anything.
Not that the remaining excess is corruption. After 2011 the payback signature is gone (see above), so what remains could be effort, nerves on the other side, or something else. The data cannot tell those apart.
Not that the purge did nothing. It may have changed who was willing, how, or on which day. It did not change how often the man on 7-7 won on the last one.
Not about the lower divisions. Wrestlers below Juryo fight seven bouts, and the schedule already pairs equal records there, so the same test does not apply.
Not about the most recent years. Since 2019 there are only a couple of dozen usable bouts a year. Both measures read lower, and the intervals are wide.
What did not work
The payback test does not single out the 7-7 bouts. It reproduces the paper's figure almost exactly. But ordinary late bouts, with nobody on the margin, show a reversal too, two-thirds as large in 1958 to 1988, while early-tournament bouts show none. So in the old record, the rematch cannot tell a repaid favour on the bubble from whatever else the end of a tournament does to the next meeting. It is reported, not used for the headline.
Hoped to use the Sumo Association as a second source for the whole history. Its site serves only the current tournament, so it became a one-tournament spot check instead.
The two main sources are probably not independent. sumo-api.com's own about page names SumoDB as the resource it set out to make easier to use. Their near-perfect agreement is therefore a parse check, not two witnesses. The independence on this page is in the measures: one is about bouts and needs a model, the other is about final records and needs none.
Record symmetry is not a usable control for measure B. Far from the line, records should mirror (10-5 as common as 5-10). They do not quite: in some periods the losing side is up to two points more common, likely from how withdrawals and visiting lower-division wrestlers fall. That asymmetry runs against the 8-7 surplus, not with it, so it is shown in the table above but not used as a gate.