Rillor tested a narrow question about crude oil. When the regulated WTI futures contract at CME moves, does the on-chain WTI perpetual on Lighter follow, and can the gap be captured after paying to trade?
The answer has two parts. The lead is real: it held on data the rules were never fitted to and survived every check we applied. A net edge after execution costs is a different matter. No execution variant we tested produced a net result that reached statistical significance.
This post summarizes the study. The full report sets out the method, the statistics and the cost measurements.
Two venues, one barrel
CME's crude oil futures contract covers 1,000 barrels, is quoted in dollars and cents per barrel with a minimum move of one cent ($10 per contract), and is delivered at Cushing, Oklahoma [1]. It trades on CME Globex from Sunday to Friday, 5:00 p.m. to 4:00 p.m. Central Time, with a 60-minute break each day [1].
Lighter runs its order book as an application-specific Layer 2 on Ethereum. A succinct proof system covers matching under price-time priority, risk management and account updates, so each match can be checked against the rules [2]. Its real-world-asset markets trade around the clock [3]. The WTI perpetual takes its price from CME futures through external oracle feeds and moves from the front month to the next in five daily steps of 20% between the 5th and 10th business day [4, 5]. When oracle data goes stale, Lighter switches to an internal price built from its own order book [5].
For most of the trading week, then, the same barrel has two prices: one formed in a regulated futures market and one formed on an on-chain order book that follows it.
Why the question matters now
On-chain crude perpetuals drew real trading in 2026. On Saturday, 28 February, the U.S. and Israel struck Iran while traditional markets were closed, and oil-linked perpetuals on Hyperliquid rose more than 5% [9]. The EIA reported that the Strait of Hormuz, which carried nearly 20% of global oil supply before the military action began that day, was effectively closed to shipping afterward, and that Brent spot reached $138 per barrel on 7 April [10].
Castle Labs compared a Hyperliquid crude perpetual with CME's April 2026 WTI contract using data from 27 February to 16 March 2026 [12]. During active trading hours the perpetual tracked CME closely. Liquidity was another story: the median volume traded within 2 basis points of the mid, per five-minute bucket, was $19 million on CME and $152,000 on the perpetual [12]. Over the first two weekends, with CME closed, the perpetual captured about 45% and then 68% of the Monday opening gap, and on both weekends it hit a ±5% price bound [12].
Those results describe size and weekend behavior. Our study asks about timing during the hours both markets trade.
What we did
We used Rillor's market data warehouse, which holds U.S. futures trade and quote history from 2018 and on-chain perpetual order-book data rebuilt from snapshots and book diffs. We defined events from moves in CME WTI futures and measured how the Lighter perpetual moved afterward, over holding periods of up to 30 minutes.
The results below are out of sample. We ran 108 tests and corrected for all of them with the Bonferroni method. Two independent implementations, one in Python and one in Rust, had to agree on every digit. A prefix-invariance check, which tests for look-ahead, confirmed that no result changed when later data was removed.
The lead is real
On held-out data, trading the Lighter perpetual in the direction CME futures had just moved earned a gross average of 14.8 basis points per event, with a t-statistic of 4.17. The result survives a Bonferroni correction across all 108 tests. The two implementations matched to the last digit, and the prefix-invariance check came back clean.
In plain terms: the regulated futures market moves first, and the on-chain perpetual follows.
The edge after costs is not proven
A gross return is measured at the price available when CME moves. A real order arrives later and pays to trade. We measured each cost directly from venue data.
| Cost component | What we measured |
|---|---|
| Fees, standard account (study period) | 0 maker, 0 taker, with about one second of added latency |
| Fees, premium account (study period) | 0.2 bps maker, 2 bps taker, without the added latency |
| Spread on the WTI perpetual | 0.8 to 1.3 bps per side, from Lighter's own trade tape |
| Funding | Immaterial for holds of up to 30 minutes |
The fee schedule forced a choice: the zero-fee account was slower, and the faster account paid fees. Taking liquidity on the lagging venue is a race decided in fractions of a second, and the size available at the stale price is small. Under those conditions, no variant's net result reached significance. Lighter's published schedule has since changed; on 9 October 2026 it lists 300 ms of taker latency for standard accounts and premium fees from 0.40 bps maker and 2.80 bps taker before staking discounts [7].
Two lessons on measurement
Estimate costs from the venue's own data. Roll's implicit spread estimator [16] put the perpetual's spread at about twice the value measured from the tape. Using another venue's top-of-book feed as a stand-in put it at about a third. Either shortcut would have put the wrong cost into the net analysis.
Measure every accounting figure against raw data. A figure derived by subtraction from a count that may have been filtered can look right and be wrong. Every cost and return in this study was measured directly.
Where the returns came from
Test-set returns for both candidate rules concentrated around 08:30 Eastern, when U.S. agencies publish the employment report, the consumer price index, GDP and personal income [20, 21, 22]. The pattern surfaced in the test data, so it needs validation on fresh data before it counts as a finding. Two other families of rules, one built on order flow and one on price gaps across trading halts, showed no effect in either instrument pair studied.
What would settle the net question
Three lines of work follow from the results. Execution as a maker on Lighter's zero-fee book removes the spread cost, at the price of adverse selection that has to be bounded with order-book data. Conditioning on the macro calendar tests the 08:30 pattern on data it was not found in. And roughly three to six times as many events would give enough statistical power to decide net significance.
This is research on how prices form across venues. It is not investment advice and not a recommendation to trade.
Read the full report
The full report, available as a PDF, covers the venues, the method, every statistic and cost measurement and the limits of the analysis, with all sources listed.