AI-generated research, checked but not peer-reviewed. It may contain errors.

E600　 KILL　 Placebo equalTested · October 2026

A zero-fee maker grid on a dollar-stablecoin pair earns nothing

A zero-fee maker grid around the USDT-USDC level returned -0.02% a year on $40,000 in the holdout half against a 5% pass line; the shuffled-path control earned +3.0%, so the real pair shows no reversion at that scale.

stablecoin peg market-making maker-grid mean-reversion zero-fee spot

Readable summary Full research record JSON recordFeedback on E600

People expect two dollar stablecoins to trade within a few hundredths of a cent of each other, because only institutions can mint and redeem them and arbitrage pulls the rate back. A trader who pays no fees could rest bids and offers a few ticks around the usual rate and collect the wobble. We tested that on one year of minute data for the USDT-USDC pair on a large US-regulated exchange with no trading fees. The grid chosen on the first half returned -0.02% a year on $40,000 in the second half, against a 5% pass line. The rate drifts instead of snapping back.

Holdout return-0.02% a yearchosen grid on $40,000, 75 cycles, after a 0.1 bp per fill stress

Pass line5% a yearthe 4% reward forgone on idle USDC was not covered either

Shuffled-path control+3.0% a yearsame grid on a path with within-day mean reversion destroyed

Selection half9 of 9 negativetotal P&L from -$191 to -$848 per configuration

01 · The idea being tested

A maker who pays no fees can earn the small oscillation of one dollar stablecoin against another.

Why people believe it

People argue that the rate between two dollar stablecoins is pinned by issuer minting and redemption, which only institutions can use and only in banking hours. Retail conversion flow then pushes the rate a few ticks away from a slowly moving level, and redemption arbitrage pulls it back, so a resting order a few ticks from that level should be filled at the extreme and closed a little later at a profit.

Why it would persist The argument, as stated before testing

People argue that mint and redeem is closed to retail accounts, so the oscillation is left to whoever can quote without paying fees.

02 · How it was tested

We took a year of one-minute prices for the USDT-USDC pair on an exchange that charges nothing on stablecoin pairs. Every hour we computed the pair's usual level as the median close of the previous 24 hours. A long grid rested a bid a set number of ticks below that level and, once filled, an offer a few ticks above the fill. A short grid did the mirror image. Orders counted as filled only when a later minute traded strictly through the price on enough volume. We tried nine settings on the first six months, picked the best by profit, and judged only that setting on the last six months. We also ran it on a price path rebuilt by shuffling each day's minute-to-minute changes, which removes any tick-level mean reversion.

Pass required

Holdout, chosen configuration: annualised return at least 5% on $40,000 after the 0.1 bp per fill stress; day-clustered 95% interval of the mean daily P&L above 0; both grids at least 0; worst cycle at least -20 bp; negative-control return below half of the real return.

Killed if

Any pass condition not met.

Costs

Zero fee on the pair; stress of 0.1 bp per fill, applied to the reported holdout figures. The 4% a year paid as a reward on USDC is the income forgone by using the capital for the grid.

Controls

Negative control: the same grid on a price path rebuilt from the holdout half's 1-minute close changes, shuffled within each day, which destroys tick-level mean reversion; it had to earn less than half of the real result. Capacity check at $100,000 per grid ($200,000 capital).

03 · What happened

Why it was rejected

The chosen configuration returned -0.02% a year against the 5% requirement, its daily interval [-0.5, +0.4] dollars included 0, and the shuffled-path control earned more (+3.0% a year) instead of less than half. The worst cycle (-6.2 bp) stayed within the -20 bp limit.

Selection half: all 9 configurations lost money in total, from -$191 to -$848 each; the configuration with k = 2 and m = 3 ticks was the least bad and was carried forward.
Holdout, chosen configuration (k = 2, m = 3): 75 completed cycles, of which 50 ended by the 24-hour timeout and 25 at the take-profit price; mean -0.03 bp per cycle, worst cycle -6.2 bp, annualised -0.02% on $40,000, day-clustered 95% interval of mean daily P&L -$0.5 to +$0.4.
No holdout configuration came near the 4% a year forgone as USDC reward on the capital.
At $100,000 per grid (volume at least $200,000 on a fill bar) the chosen configuration returned -0.18% a year on $200,000.
Negative control: the same grid on a path with the holdout's minute-to-minute changes shuffled within each day earned +3.0% a year, far more than the real path's -0.02%.
The pair's 24-hour median level moved between 0.9981 and 1.0005 over the year. Of the 75 holdout cycles, 50 never reached their take-profit price within 24 hours.

Holdout returns against the 5% pass line

Annualised return on capital of the grid configuration chosen on the first half, judged on the last half. The $40,000 case returned -0.02% and the $100,000-per-grid case -0.18%. The 5% pass line and the 4% reward forgone on idle USDC are far above both bars.

Chart data (% a year)

	Chosen configuration, holdout
$20,000 per grid ($40,000 capital)	-0.02
$100,000 per grid ($200,000 capital)	-0.18

The shuffled path earns more than the real one

Real holdout pathShuffled within-day path

The same chosen grid on the real holdout path and on a path whose minute-to-minute changes were shuffled within each day. A grid that harvests tick-level mean reversion would earn less on the shuffled path; it earned +3.0% against -0.02%, so the real pair does not revert at the tick level.

Chart data (% a year)

	Real holdout path	Shuffled within-day path
Chosen configuration, $40,000	-0.02	3

Two of three cycles ended by timeout

How the 75 holdout cycles of the chosen configuration ended: 50 were closed by the 24-hour timeout, 25 reached the take-profit price. A timeout closes the position one tick worse than that bar's close.

Chart data (cycles)

	Cycles
Closed at take-profit	25
Closed by 24-hour timeout	50

04 · What it means

Why it failed Interpretation · not established

The rate between the two coins behaves like a level that drifts, not like a price that oscillates around a fixed one. A resting order is filled strictly through its price mostly when the level itself is moving, so the position then waits on the wrong side and ends by timeout. The shuffled-path result is consistent with this: removing the real ordering of changes made the grid earn more, which means real tick-level reversion was not what the real path contained. This is an interpretation; the test did not measure queue position or order flow.

Lesson

A quote at zero fee only earns when the price it is resting against comes back; whether it comes back is a property of the price path, and a pegged asset can still drift in level while staying inside a very narrow band. A negative control that destroys the mechanism is a cheap check: if it earns as much or more than the real data, the real result was not the mechanism.

Scope of this result

Covers one pair of dollar stablecoins on one large US-regulated exchange with no fees, one year of 1-minute candles, nine grid configurations, conservative trade-through fills without a queue model, $20,000 and $100,000 per grid. It does not cover fills as a real order book would grant them, other venues, or other stablecoins.

A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.

1. Test design

Pre-registered

Yes: rule and criteria fixed before the test data was examined

Data

1-minute candles of the USDT-USDC spot pair, 221,176 traded minutes, median daily volume $2.9 million — a large US-regulated exchange, public market-data API; 1 minute; 2025-10..2026-10; public

Universe

One pair (USDT-USDC) on one venue with no trading fees on stable pairs. Only minutes whose traded volume was at least twice the order size were eligible. The 24-hour median level ranged from 0.9981 to 1.0005.

Signal

Level c = median 1-minute close of the previous 24 hours, recomputed every hour and rounded to the tick (0.0001). Long grid: when flat, rest a bid at c - k ticks. Short grid: when flat, rest an offer at c + k ticks. Configurations: k in {0, 1, 2} ticks times m in {1, 2, 3} ticks, nine in all.

Execution

A resting buy at price p is filled only if a later 1-minute bar trades strictly below p (low < p) and the bar's volume is at least 2 times the order size; sells mirror (high > q). No queue model. Once filled, the long grid rests an offer at the fill plus m ticks (the short grid a bid at the fill minus m ticks). A position not closed within 24 hours is closed at that bar's close, one tick worse. Size $20,000 per grid, $40,000 of capital. The minimum-volume multiple was changed from 10 to 2 before the year of candles was read (only the last 6 hours had been seen), because the pair trades only about $3,000 to $4,000 a minute on average.

Costs

Zero fee on the pair; stress of 0.1 bp per fill, applied to the reported holdout figures. The 4% a year paid as a reward on USDC is the income forgone by using the capital for the grid.

Controls

Sample split

First six months for choosing the configuration by total profit; last six months for judging it. One configuration was carried into the holdout.

Pass criterion

Kill criterion

Any pass condition not met.

Power

The holdout produced 75 completed cycles. The daily-P&L interval is about +/-$0.5 on $40,000, while a 5% annual return is about $5.5 per day, so the test could have seen a result at the pass line with room to spare. The nine configurations are chosen on one half, so results across them are not independent.

2. Result

Observations

Instruments

Net mean

-0.03 bp

95% interval

Not reported

Median

—

Worst

-6.2 bp

Finding 1

Selection half: all 9 configurations lost money in total, from -$191 to -$848 each; the configuration with k = 2 and m = 3 ticks was the least bad and was carried forward.

Finding 2

Holdout, chosen configuration (k = 2, m = 3): 75 completed cycles, of which 50 ended by the 24-hour timeout and 25 at the take-profit price; mean -0.03 bp per cycle, worst cycle -6.2 bp, annualised -0.02% on $40,000, day-clustered 95% interval of mean daily P&L -$0.5 to +$0.4.

Finding 3

No holdout configuration came near the 4% a year forgone as USDC reward on the capital.

Finding 4

At $100,000 per grid (volume at least $200,000 on a fill bar) the chosen configuration returned -0.18% a year on $200,000.

Finding 5

Negative control: the same grid on a path with the holdout's minute-to-minute changes shuffled within each day earned +3.0% a year, far more than the real path's -0.02%.

Finding 6

The pair's 24-hour median level moved between 0.9981 and 1.0005 over the year. Of the 75 holdout cycles, 50 never reached their take-profit price within 24 hours.

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Download 1-minute candles of the USDT-USDC spot pair for one year from the exchange's public market-data API. 2) Compute the level as the median 1-minute close of the previous 24 hours, recomputed hourly and rounded to 0.0001. 3) Simulate a long grid (bid at level minus k ticks, offer at fill plus m ticks) and a short grid (mirror), with fills only when a later bar trades strictly through the price on volume of at least twice the order size, and a 24-hour timeout exit at the close one tick worse. Use $20,000 per grid, 0.1 bp stress per fill. 4) Run the nine combinations of k in {0, 1, 2} and m in {1, 2, 3} on the first half and pick the best by total P&L. 5) Evaluate that one on the second half: annualised return on $40,000, bootstrap of daily P&L by day, worst cycle. 6) Repeat on a path rebuilt from the holdout's 1-minute close changes shuffled within each day, and at $100,000 per grid.

Grade

COMPLETE