The Lab

Where the idea comes from

• Tokenised gold and stocks also trade on chain, in pools on decentralised exchanges, beside the crypto exchanges' order books and perps.
  • Gold: PAXG and XAUT, on Ethereum.
  • Stocks: the xStocks tokens (AAPLx for Apple, TSLAx for Tesla, SPYx for the S&P 500 fund and others), on Solana.
• A pool's depth is set by how its money is spread across prices. Uniswap's third version (Hayden Adams and others, 2021) lets each provider place money in a price range. A pool is deep near the price only if providers put money there.
• Which market leads is a question Joel Hasbrouck (1995) put for a stock traded in many places: the price moves first where most of the informed trading is, and the others follow.
• Prices that update rarely look as if they follow. Andrew Lo and Craig MacKinlay (1990) showed that a thinly traded price catches up late, so it trails the busy one even with no one acting on it.

How we tested

• The tokens: PAXG and XAUT; nine xStocks (AAPLx, AMZNx, COINx, CRCLx, GOOGLx, HOODx, METAx, NVDAx, TSLAx), plus SPYx and QQQx.
• The window: 1 August to 19 October 2026, 79 days.
• On chain: every pool GeckoTerminal lists for the token, on any network.
  • Volume: the daily candles of each token's five largest pools. Smaller pools were counted from their last 24 hours times 79 days, an estimate.
• On the exchanges: Bybit's spot token and its perp on the same asset, and Binance's and OKX's where they list one.
• Depth, on 21 October 2026, while the US stock market was open:
  • 10 readings, 60 seconds apart, of the cost of a market buy and sell;
  • on Solana, Jupiter's quote (its best route over all pools);
  • on Ethereum, Uniswap's own quote contract, read through a public node (a read only, no wallet);
  • on the exchanges, a walk through Bybit's order books.
• Lead or follow: hourly prices of each token's largest pool against its Bybit perp.
  • The control: the perp's days dealt in random order, 1,000 times.
  • The bar, fixed before the run: on chain leads if the result is above 95% of the shuffles, and follows if below 5% of them.

Added after the first run. GeckoTerminal's search also returns impostor tokens under the same name. On Solana, some "PAXG" and "XAUT" pools claimed reserves of billions of dollars. We counted only the issuers' own tokens: gold's Ethereum contracts, and each xStock's Solana address. The impostors had carried little volume, so the shares barely moved (PAXG on chain 6.3% → 5.4%), but they had pushed PAXG and XAUT out of the lead-or-follow test. With them removed, PAXG is in it. XAUT's largest pool had too few hourly prices (21), so it stayed out.

The maths, step by step

1. The cost beyond the middle. Buy and sell 100 USD; the middle of the two prices is the reference. A 100,000 USD buy at an average of 100.6 against a middle of 100 cost 0.6%. On chain this includes the pool's fee.

2. A correlation an hour apart. Take each hour's move on chain and the perp's move in the next hour. If on chain leads, the two line up: the correlation is above 0.

• LEAD = the correlation of chain at hour t with the perp at t + 1.
• FOLLOW = the correlation of the perp at hour t with chain at t + 1.
• The test is LEAD minus FOLLOW, averaged over the tokens.

What we found

1. Most of the trading happened on the perps. Shares of each token's volume, 1 August to 19 October 2026:

• The perps took 66% to 98% of the volume for every token but one. On chain carried 1% to 12%.
• SPYx was the exception: two thirds of its volume was on chain, 758 million USD over 79 days, on pools holding about 6 million USD. That is each dollar in the pools traded about 1.6 times a day. Whether that is people trading or bots trading among themselves, these figures don't show.
• Each stock token traded mostly on Solana. The gold tokens' genuine pools were all on Ethereum.

2. Both kept trading while the stock market was shut. The US market's regular session covers about a fifth of a week's hours. The share of volume in the other hours:

• About 40% to 70% of on-chain volume came while the market was shut; 43% to 77% on the perps. With four fifths of the hours shut, trading was still busier, hour for hour, while the market was open.

3. On chain was much thinner near the price. The cost of a market order beyond the middle, the median of 10 readings:

On chain: the best route at that moment. The perp: Bybit's book. "Times" is on chain's cost over the perp's, from the unrounded costs.

• At 100,000 USD, on chain cost 5 to 32 times what the perp did: 0.28% (SPYx) to 2.5% (HOODx), against 0.02% to 0.16% on the perps.
• At 1,000 USD the costs were small: 0.01% to 0.22% on chain, mostly the fees, against 0.03% or less on the perps.
• Bybit's spot tokens sat in between. In at least one reading, the books of CRCLX and HOODX were nearly empty on one side at 100,000 (costs of 26% and 59%); their medians were 0.64% and 1.67%.

4. On-chain prices followed the perps.

• LEAD minus FOLLOW was −0.041, below all 1,000 shuffles: on chain follows, by the bar set before the run.
• The effect is small. A correlation of 0.05 means the perp's hour tells little about the pool's next hour.
• It fits Lo and MacKinlay's stale prices: a pool that trades rarely catches up with the busy market an hour later. It doesn't show that anyone could profit from the delay. The pool's fee and spread, 0.01% to 0.1% for a small order, are larger than most hourly catch-ups.

Limits

• 79 days, August to October 2026, a calm stretch for gold and the big stocks.
• Smaller pools' volume is an estimate, their last 24 hours times 79 days. It is up to half of a token's on-chain total (GOOGLx 51%).
• Volume figures are as each source reports them. Some may be bots trading with themselves, on chain or off; we can't separate it.
• The depth is one hour's readings on a quiet day with the market open. A crash, or a weekend, could show less.
• Only the issuers' own token addresses count. Copies bridged to other chains were left out (for example, a token named XAUT on BNB Chain), so on-chain shares could be a little higher.
• The perp's hourly price is Bybit's alone. Binance's or OKX's could lead by a little more or less.

Update, 21 October 2026: the depth table

Added after this page was published. The text above is left as it was, except the changes this lists.

• The depth table gained the 1,000 USD columns (on chain and the perp) and a "times" column, on chain's cost over the perp's at 100,000 USD, from the unrounded costs. The 10,000 USD columns were dropped.
• Three cells moved by rounding, as the table is now built from the unrounded costs:
  • PAXG, the perp at 100,000: 0.05% → 0.04%;
  • TSLAx, the perp at 100,000: 0.07% → 0.06%;
  • NVDAx, on chain at 100,000: 0.30% → 0.29%.
• One line reworded to match the table. Before: "At 1,000 USD the gap was small: 0.01% to 0.22% on chain, mostly the fees."
• The findings and the text's figures are unchanged.

Credits

• Hayden Adams, Noah Zinsmeister, Moody Salem, River Keefer and Dan Robinson, "Uniswap v3 Core", 2021: how a pool's depth is set.
• Joel Hasbrouck, "One Security, Many Markets: Determining the Contributions to Price Discovery", Journal of Finance, 1995: the question of which market leads.
• Andrew W. Lo and A. Craig MacKinlay, "An Econometric Analysis of Nonsynchronous Trading", Journal of Econometrics, 1990: why a stale price seems to follow.
• Paxos (PAXG), TG Commodities, Tether's gold company (XAUT), and xStocks: the tokens' own addresses.
• Our own: the volume shares, the depth comparison and the lead-or-follow test with its shuffles.
• Data: GeckoTerminal's public API; Jupiter's public quote API; a public Ethereum node (read only); Bybit's public market data; Binance's public archive; OKX's public market data; Yahoo's public prices for the US market's trading days, as the reference only.
• PerpGremlin and Claude, Tokenised gold and stocks on chain: where they trade, TerminalGremlin, 2026: the tests, the code and the words.