The Lab

Where the idea comes from

• Grid trading is long-standing market practice, with no known originator. Running a long grid and a short grid on the same coin together is sold by exchanges as a "neutral" grid.
• A grid earns only if prices come back to where they were: our theory page "Mean reversion: does what drifts away come back?".
• The carry it is measured against is the hedged trade that Maik Schmeling, Andreas Schrimpf and Karamfil Todorov measured in crypto (2023), on our page "The same long, held two ways: plain coins, or coins plus a hedge that earns funding".
• Not the same as pairs trading. Our page "Pairs trading: does the gap between two coins come back?" trades the gap between two different coins. This page is two grids on the one coin.
• What this page adds to our grid pages: those test one grid alone ("Grid bots on 14 coins: what 30 months of prices say", "Grid bots: more rungs or fewer?"). Here the two grids net against each other in one account, and share one liquidation line.

What the pair is

• The long grid buys on each rung under the starting price, and sells each buy one rung up.
• The short grid sells on each rung over the starting price, and buys each back one rung down.
• At the start both hold nothing. Under the start only the long grid holds coins; over it only the short grid holds a short.
• So the pair is long when the price is low and short when it is high. It earns one gap on every swing back. It loses when the price runs away from the start and stays away.
• On the exchange: the two grids need hedge mode, which keeps a long and a short open on the same coin at once. In one-way mode their orders would cancel each other out. Our own engine runs the pair this way, one bot for each side. A demo pair ran for four days in October 2026: too short to judge by.

How we tested

• Bitcoin, every minute from April 2020 to September 2026: 78 months. This takes in the 2020–21 rise, the 2022 fall and the years since.
• Both grids set fresh each month on the same rungs, around the month's first price:
  • a range of 15% under the start to the same distance over it, or 30%;
  • 8, 20 or 50 rungs, evenly spaced in percent;
  • closed at the month's last price, and set again around the new price.
• Leverage: each grid gets half the money, at 2x, 3x, 5x, 7x or 10x. Only one grid holds much at a time, so the account's exposure is at most half that: 5x a grid is at most 2.5 times the money.
• One cross-margin account. It is liquidated at the first minute where the two grids together lose all the money, allowing for the exchange's maintenance margin (0.5% of each position). That month counts as −100%, and the account is gone.
• Real costs: 0.02% on every grid order, 0.055% to close at the month's end. Each grid pays or receives the actual funding on what it holds.
• The two to beat, on the same months:
  • Hold: Bitcoin bought at the start (0.1% fee) and held, no leverage.
  • Carry: Bitcoin held and the same dollars shorted on Bybit's coin-settled perpetual, collecting each funding payment, with full fees in and out. It is the same book as on our carry page: from June 2022 to September 2026 both give +27.58%.
• Fair tests, fixed before the run:
  • every setting is shown;
  • the first 39 months pick the best setting, and the last 39 judge it;
  • each month is also run 10 times with its hours shuffled into random order: the same moves, no pattern. Does the real order pay more?

The maths, step by step

1. What one swing earns. The 30% range is 0.7 to 1.43 times the start. That is a factor of 2.04, split into 19 gaps of 3.8% each. A lot is half the money ÷ 20 rungs = 2.5% of the money. One round trip earns 3.8% × 2.5% = 0.095% of the money, less 0.001% in fees.

2. A month of swings. About 20 trips a month gives 1.9% of the money at 1x per grid.

3. What a trend costs. Say the price ends the month 15% up. The short grid has sold on the 4 rungs it passed, about 2%, 6%, 10% and 14% over the start. At the month's end each is under water by about 13%, 9%, 5% and 1%. That is (13 + 9 + 5 + 1)% × 2.5% = 0.70% of the money.

4. Leverage multiplies both. At 5x a grid each figure is 5 times bigger: about 9.5% from the swings, and 3.5% lost to that trend.

5. Over months, the swing in returns costs too. A month of +50% then one of −50% leaves 0.75 of the money, not 1. The more the months swing, the less the same average month compounds to.

What we found

1. Neither beat holding Bitcoin over the whole time, and none beat the carry.

Worst drop: from a month-end high.

• Best of all 30: the 30% range, 20 rungs, 5x a grid: ×1.46, under a quarter of the carry's gain, with a drop of 39% against the carry's 2%.
• The 15% range lost money at every leverage and rung count. Every 10x book on it was liquidated, and so were 7x with 20 or 50 rungs.
• The 30% range was never liquidated. But at 10x with 20 rungs it still fell 54% in its worst month.
• More leverage did not mean more money. On the 30% range with 20 rungs, 5x and 7x ended highest, and 10x ended lower: its swings cost more than its extra average (maths, step 5).

2. The swings paid, then the trends took it back. Per month at 1x a grid:

• The swings earned well, but what was left open at each month's end took nearly all of it back.
• The long grid made money and the short grid lost (30% range: +0.37% and −0.19% a month). Bitcoin rose over these years, so the short side kept being caught out.
• Funding helped a little. Funding is usually paid to shorts, and on balance the pair received it.

3. The real order of prices was no better than random order, and on the narrow range worse. Each month was run 10 times with its hours shuffled:

• 30% range: the real months beat their shuffled twins 6, 7 and 4 times in 10 (8, 20 and 50 rungs), around the 5 in 10 that luck gives.
• 15% range, 20 and 50 rungs: the real months did worse than all 10 shuffles. Bitcoin's real hours run in streaks more than shuffled ones do, and a streak carries the price through a narrow range and leaves it outside.

The pair caught no pattern in how Bitcoin moves that paid.

4. Over a year, the pair sometimes wins; over three, almost never.

It beat holding in the years when Bitcoin fell or went sideways, and in 2021, when it swung hard both ways:

30% range, 20 rungs.

5. The pick from the first half lost in the second. The first 39 months picked the 30% range, 8 rungs, 7x a grid (×1.38 there). In the last 39 it ended at ×0.89, 21st of 30 settings. Over the same 39 months, holding gave ×2.74 and the carry ×1.28.

What it means

• Two opposing grids are one bet: that the price comes back. They earn from swings and give it back on trends. Over these years, Bitcoin trended more than it swung back.
• Against holding, the pair won in falling or flat years and in 2021's big swings, but holding's gain in the rising years was far larger.
• Against the carry, the pair earned less, with far bigger drops. The carry's hedge never runs into a trend, because its short exactly matches its coins.
• Leverage didn't rescue it. Past 5x a grid it lost more to the bad months than it gained, and the narrow range was liquidated.
• No setting was reliable: the first half's best lost money in the second half.