Margin Call from Seoul

BTC spent the week going nowhere fast, after Monday flush a slow three-day repair ensued. A late bid on Thursday nearly got it back in the black, but it ended down 0.5% since last Friday, it’s still on for its best month since April and its best July in four years. Elsewhere there was plenty of dispersion, BNB led the altcoins, while ETH and DOGE also ground higher.

The week had two catalysts, Monday it was Korea, as the KOSPI unwind that runs through this issue's spotlight bled into everything risk-adjacent and handed crypto its worst day of the week.

The predictable catalyst of the week was the Fed, and Citadel and others caused some increased uncertainty around the meeting, as fears of a hike spiked. However, rates remained unch but a divided committee, with three dissents, made the headlines.

Equities and gold closed the week telling a calmer story. The volatility lived inside the week, not across it.

Price and flow disagreed all week, and looking at key indicators across both spot and perps makes it clear. BTC maintained steady net buying and still closed red, while everything that fell got sold with conviction. The week's best performer went up on flow of roughly zero, nobody bought it.

The biggest day of the week in terms of flows was not the Fed, but Monday’s APAC inspired selling. Wednesday was a microstructure event more than a price event though. Spreads widened in both crypto and traditional markets, with GLD and QQQ doubling (or more) their quoted cost and BTC's average spread roughly tripling, though the two moves lived in different places.

The ETF widening ran through the whole book, with the typical quote getting worse, while BTC’s didn’t. BTC’s median quote never left the minimum tick, and the damage sat entirely in the tails, in the handful of prints around the decision when the book stretched to many times its resting width. The rest of the majors held their tick, and depth thinned everywhere.

One caveat on reading the chart across rows. The ETF lines are each exchange's own book rather than the consolidated tape, so the quoted cost sits wider than what a trader actually pays. The multiples are the story here, not the widths.

Through the window itself, BTC was bid for the full three hours and kept what it gathered. ETH chased the statement and gave a third of it back once the headlines landed from Warsh’s speech. HYPE never caught a bid, remaining flat until the press conference.

One share of SK Hynix printed 30% below the prior close in the Korean pre-market on Tuesday, an hour before Seoul opened when markets were thin. As a result the related perp on Trade.xyz fell over 17% in sixty seconds, liquidating longs for around $60mn.

Trade.xyz, the perps venue on Hyperliquid specialising in real-world assets, and one of them tracks SK Hynix. The contract tracks the underlying asset’s moves on the KRX, with prices being fed via an oracle which itself receives data from third parties. That all worked as advertised, albeit with some flaws in the architecture.

The catch is the hours, because the perp never closes while the Seoul market is only open for six and a half hours a day during the week. This leaves the oracle quoting a live price for most of the week on a stock that is not actually trading.

This is exactly where the risk lies, and where the anomalous print led to liquidations. At eight in the morning Seoul time on Tuesday the 28th (11pm UTC on the 27th), a single share reportedly printed in the Korean pre-market roughly 30% below the prior close. The price was back where it started a few minutes later, but the damage had already happened on Trade.xyz.

For scale, a normal pre-open hour for this stock has a range of about two and a half percent, and this one printed 23%, multiples of any other morning session in the past month (and its been a volatile month on the KRX). The earnings that genuinely cratered the stock were still a day away, so this was not anticipation, because there was nothing to anticipate. When earnings did disappoint the market didn’t even hit double digit range.

Data can’t tell us the intent behind a trade, but it can (and did) tell us how robust oracle infrastructure is onchain. Seven venues (mostly CEXes) run perps on SK Hynix, and every one of them ingested the same anomalous print. Six of them landed together at around 6% down. Trade.xyz was alone with its 17.8% drop.

So are the other venues wrong, or did Trade.xyz’s reference price fail? If the venue refunding users didn’t tip you off, this should. The same anomalous print hit all seven venues, but only Trade.xyz’s contract fell double digits, then the question is not who was right, but whose reference methodology was least robust. Trade.xyz’s pricing logic appears to have been materially more fragile, because a single thin pre-open trade should not have been able to dominate the mark so easily.

That does not mean the market violated some grand pricing theory; it means the reference mechanism was vulnerable to a bad print. The problem is not that the oracle ignored market information, but that it appears to have let a single anomalous trade steer the reference price too far.

Trade.xyz's own protection tells the uncomfortable story. SKHX carries a plus or minus 10% discovery bound with one reset, which compounds to a floor roughly 19% below reference, and the low printed just inside it. From the venue’s point of view the oracle worked, it turned a 30% down input into a near 18% output. But in the context of competing venues and $60mn in liquidations, there’s working as intended and working as traders expect. It did not satisfy the latter, hence refunds.

Liquidations trigger continuously on the way down rather than at the floor, so halving the damage still liquidated everyone caught in the first half.

One share should not move a market carrying half a billion dollars of open interest. Equally, building a derivative on a high volatility KRX stock makes this difficult, but not impossible. There are some rules venues should build in to protect users, especially as AI-linked vol looks set to stay. One idea could be keeping prices bounded between bid and ask spreads overnight, and for any prints outside of this range the weight of the input is lowered the further from the bound it is.

Crypto's earnings season opened this week, and it was as expected. Coinbase and Strategy both reported Wednesday, with the former impacted heavily by lower volumes, though prediction markets were the bright spot in Coinbase's deck.

Circle reports Wednesday the 5th, the first read on stablecoin economics in a quarter where supply went sideways, although the higher for longer rates environment should benefit stablecoin issuers.

Something to watch in terms of macro and rates outlook is next Friday’s jobs report, and it arrives with three hike dissents freshly on the record. CPI follows the week after. Congress rises for the summer break on the 10th of August, which closes the CLARITY Act's window and makes the odds of it passing before midterms essentially null. Jackson Hole caps the month, Warsh's first as Chair, and with the theme being payments there’s bound to be some extra crypto interest.