Research · 30 July 2026
Findings from 1,028 auction lots across fourteen sales and five houses, March to July 2026. Every one captured from the house's own published results.
Only two of the sales I captured published estimates at all. Across those two — 219 lots, two different sports — 63% closed below the low estimate. The median lot made 0.75× the bottom of its range.
Sotheby's Summer Marquee: 61% below low, median 0.80×. The Beautiful Game: 72% below low, median 0.60×.
These aren't dishonest numbers. They're unchecked ones. Nobody has ever gone back and graded an auction estimate, so nothing corrects.
This is the cleanest natural experiment in the dataset, and I didn't design it — the auction did.
Sotheby's Finals Premier sale contained ten game-issued jerseys. All from the same game. All matched to the same locker room image. All bench players. The only variable was which team won the title.
| Team | Lots | Median |
|---|---|---|
| Knicks — won | 5 | $6,400 |
| Spurs — lost | 5 | $1,408 |
The distributions don't overlap. The cheapest Knicks jersey beat the most expensive Spurs one.
The nets from the same series say it too: the Game 5 championship-clinching net made $179,200, the Game 1 net $89,600. Exactly two times.
Nikola Jokić's Christmas Day jersey sold twice this year.
| Sale | Result |
|---|---|
| Christmas Premier | $177,800 |
| Christmas in July — weekly sale | $12,160 |
Same player. Same event. Same jersey type. And the cheap one carried a 39-point triple-double.
It happens across the board. Christmas Premier ran 8.3× the weekly on median. All-Star Premier ran 4.5× the All-Star weekly — same weekend, same event. The Finals sales decayed across three catalogues in a fortnight.
And it changes what a category is worth. Shorts were bottom-tier at Finals Part III, median $512. At the All-Star weekly they took the top four slots — above every jersey in the sale — because the jerseys had been held back for the Premier catalogue.
Two sales in this dataset came direct from the athlete: the Stephen Curry Collection in April and the Scottie Pippen Collection in March.
Roughly three quarters of the Curry lots beat their estimate — the exact inverse of everything else. The Pippen sale ran a median of $28,160, about 44× the median of a weekly sale.
Then this. A Kobe 4 Protro “Gold Medal”, worn by Curry, sold in the Curry Collection for $70,400. Ten weeks later the same shoe sold at Goldin for $48,800.
Down 31%. Nothing about the shoe changed. What changed is that the event was over.
One more thing from Pippen's sale worth knowing: his two highest lots weren't his. Larry Bird's Dream Team jersey made $896,000 and Michael Jordan's Dream Team sneakers made $640,000, both from Pippen's own collection. His best own-worn lot was the Flu Game jersey at $486,400.
Across 73 basketball game-used lots at Heritage in 2026, photo-matched lots ran a median of $9,760. Unmatched, $1,738.
It's confounded — better lots get matched — but the size of it is hard to argue with. And there's a cleaner version from SCP in July: the same pair of Jordan's 1998 Finals shoes, split by half. The photo-matched half sold for $2.2 million. The half where the match failed sold for $102,000.
Though it isn't the match alone. Two Last Dance Air Jordan XIIIs, both worn, both signed: the one attributed to the playoff run with no match made $204,000. The one conclusively matched to an ordinary January game made $84,000.
The match isn't what pays. The match to a moment is what pays.
In one sale in May, two lots were listed word for word identically. Same player, same shoe, same player exclusive, same charity.
One sold for $512. The other for $2,560.
Five times apart, for objects nobody could tell apart. That's the limit on how accurate any prediction in this market can be. Some of what looks like a bad call is just which minute a lot closed in.
Anyone selling you a valuation model that doesn't acknowledge that is selling you something.
I published a version of the first finding built on truncated lot titles rather than the actual catalogue entries. Two of my four examples turned out to be something other than what I'd called them — one was a warm-up jacket, one had an “apparent” match rather than a conclusive one — and a lot that belonged in the sample was never looked for.
I withdrew it. The retraction is here, and the rule that came out of it is simple enough that I'm embarrassed it needed writing down: capture the full results page before analysing anything, and check the lot count on the page against what you have.
I also had a finding about counting milestones — the 10,000th point, the 1,000th game — underperforming badly. It looked like the result of the month on one sale. It failed to replicate on the next. It stays out until it doesn't.
The sample leans heavily basketball. Five houses, not fifty. Only two of the fourteen sales published estimates, so the accuracy finding — the strongest one — rests on 219 lots. Live-floor sales resist this method entirely, because the visible bid stops meaning much once a room is involved.
I'd rather say that here than have someone else say it for me.