▲ Work in progress. This site is built and researched in the open: pages, numbers and findings may change or be wrong, and the record has gaps. Not advice. More
The Lab
Ideas traders use, put to the record
← the Lab · grids · write-up
#20 Grid bots, followed up: range width, the stop, and four years of ETH
Past prices; no real orders were placed. Every figure is a share of the money put in, or a count.
Last checked: 16 October 2026.
Where the idea comes from
- Grid trading is long-standing market practice, with no known originator.
- A grid is a rebalancing machine: it buys after each step down and sells after each step up, which is the idea behind "Shannon's demon". See our theory page "Shannon's demon: making money from a price that goes nowhere".
- It earns only if prices come back to where they were. That is mean reversion, on our theory page "Mean reversion: does what drifts away come back?".
What's new against the published grid study
The grid study on 14 coins (April 2024 to September 2026) found that a long grid follows its coin, that wider ranges lose less, and that the stop cost money on average. These follow-ups add four things:
- A longer window. Four years of ETH, BTC and SOL (October 2022 to September 2026), not 30 months.
- The stop pulled apart from the range. In the grid study a wider range also put the stop further down, so the two couldn't be told apart. Here they are crossed.
- The stop's firing rule. Does waiting for an hourly close, instead of 5 minutes, stop it selling wicks?
- A wick is the thin line above or below a candle's body: how far the price went during the candle before coming back. For example, a candle that opens at 100, dips to 95 and closes at 99 has a wick down to 95.
- A correction. On ETH's four years, ranges of 25–30% either side looked positive in both halves. That was seen after the fact, and it didn't hold on the 14-coin window. A wider range is less risk, not shown to earn.
How we tested
- The grid: a long grid on the perpetual, 8 rungs (6, 10 and 12 were tried too), 2x leverage. It starts fresh at the first price of each month and closes at the month's end.
- The range: either an uneven shape around the price (the bottom 17% under, the top 15% over), or an even 10%, 15%, 20%, 25% or 30% either side.
- The stop: none, or 3%, 7.1%, 12% or 18% under the bottom rung. It fires one of four ways:
- the mark price under it for 5 minutes;
- The mark price is the price a futures exchange uses to value positions and to decide liquidations. It is built from several exchanges' spot prices, so one wild trade can't liquidate anyone. For example, a sell that knocks the traded price down 3% for a second may barely move the mark.
- a 5-minute close under it;
- a 1-hour close under it;
- a 4-hour close under it.
- the mark price under it for 5 minutes;
- Costs: the engine's own planner and backtest, 0.02% for a resting order, 0.055% for the stop's market sale, and the funding the exchange charged at the time.
- Prices: Binance's hourly candles and one-minute mark prices, each file checked against its published checksum.
- Fairness:
- settings are chosen on the first two years and judged on the last two;
- every month is also run with its hours shuffled into random order;
- every setting is printed, not only the best.
The maths, step by step
1. The shape. With ETH at 2,530:
- a bottom at 2,100 is 17% under;
- a top at 2,900 is 15% over;
- a stop 7.1% under the bottom is 2,100 × (1 − 0.071) = 1,951.
8 rungs from 2,100 to 2,900 have 7 equal gaps in percent: the 7th root of 2,900 ÷ 2,100 = 1.381 is 1.047, a gap of 4.7%.
2. Why rung count barely matters. A round trip pays a resting fee twice: 0.02% × 2 = 0.04%, against a 4.7% gap. At these gaps the fee is under 1% of what a trip earns. More rungs give more trips, but each earns less, and the total comes out about the same.
3. The average, and how sure it is. A mean month is the plain average. A median month is the middle one when they are lined up.
- Example: the uneven shape on ETH made a median of +3.1% a month but a mean of −0.75%. Most months won a little, and a few lost a lot.
- The standard error says how far the mean could be off by luck. It is the months' spread divided by √(number of months).
- Here the spread was about 10.4% a month over 48 months: 10.4 ÷ √48 = 1.5%.
- A mean of −0.75% ± 1.5% can't be told from zero.
4. Beating shuffled months. Each month was rerun 20 times with its hours in random order: the same moves, but none of the habits of real prices.
- If real prices had nothing special, the real month would be just one more among 21.
- So "beat 18 of 20 shuffles" happens by luck 3 times in 21 (14%), and "beat 19 or 20" 2 times in 21 (10%).
- "16 to 19 of 20 shuffles did better" means the real month sat near the bottom.
What we found
1. Four years of the uneven shape on ETH: no better than zero.
| round trips | 5 a month |
| left the range | above in 42% of months, below in 27% |
| per month | median +3.1%, mean −0.75% (± 1.5%) |
| losing months | 29% |
| worst month | −32% (the August 2024 crash) |
| stop fired | 15% of months; 39% of three-month stretches left alone |
| funding paid | about 0.2% a month |
Many small wins and a few large losses cancelled out over four years.
2. The width of the range matters more than the stop.
Mean month on ETH, first two years / last two, averaged over the other settings:
| range | first | second |
|---|---|---|
| uneven shape / 15% | +1.6% / +1.2% | −1.7% / −1.5% |
| 20% | +2.1% | 0.0% |
| 25% / 30% | +1.4% / +2.0% | +0.3% / +0.7% |
With the stop held still, the uneven shape and 15% lost in the last two years, and 25–30% did not. That held with no stop too.
3. The stop's distance is a trade, not a gain.
| stop under the bottom | first | second | fired in 3 months left alone | worst 3 months |
|---|---|---|---|---|
| 3% | +1.5% | −0.8% | 38% | −30% |
| 7.1% | +1.5% | −0.6% | 34% | −36% |
| 12% | +1.6% | −0.3% | 28% | −41% |
| 18% | +2.0% | −0.1% | 21% | −49% |
Further down, the stop fires less and the average month is a little better, but the worst three months get much worse. It is a choice of risk, not something the data can optimise.
4. The firing rule barely matters, and its one gain is one event.
- Waiting for a 1-hour close made +0.1% a month instead of −0.75%. Nearly all of that is one crash, in April 2025, that wicked under the stop and came back within the hour.
- Of the other six times the stop fired, the hourly rule did better twice (by 7.5 and 0.7 points), tied once, and did worse three times (by 0.3 to 1.3 points). It sells later in a real fall.
5. Choosing on the first half didn't pick the winners.
- The first half's best on ETH (15%, 6 rungs) ranked 14th of 24 in the second.
- The best overall when the stop was crossed in (20%, no stop) made +0.5% a month in the second half, 21st of 85. It beat 18 of 20 shuffles.
- The uneven shape with its 7.1% stop on the 5-minute rule came last of 85 in the second half.
6. On BTC and SOL the wide ranges looked positive too, in this window only. Mean month, first two years / last two:
| range | BTC | SOL | ETH |
|---|---|---|---|
| uneven shape | +2.1% / −0.4% | −0.4% / +0.2% | +1.2% / −2.7% |
| 15% | +1.4% / −0.2% | −0.3% / −1.7% | +1.7% / −2.1% |
| 20% | +2.2% / +0.1% | +0.5% / +0.5% | +1.7% / −0.2% |
| 25% | +2.0% / +0.4% | +1.1% / +0.9% | +0.6% / +0.1% |
| 30% | +1.9% / +0.1% | +2.0% / +0.8% | +2.1% / +0.5% |
- The second-half gains are small, 0.1–0.9% a month, each within one standard error of nothing.
- On the 14-coin window (April 2024 to September 2026) the same settings were positive in both halves on only 1 of 14 coins at 25%, and 2 at 30%.
- And 25–30% was picked for this test from ETH's two halves together, not from its first half.
7. Against shuffled months, real prices didn't come back to their range more than random ones.
- On ETH, the uneven shape and narrow ranges earned less than the shuffled months: 16 to 19 of 20 shuffles did better. At that scale ETH trended more than it came back.
- Only ETH's 30% range beat the shuffles (19 or 20 of 20). By luck, any one setting does that about 1 time in 10, so among twelve settings about one would.
- SOL earned less than its shuffles at every width.
8. BTC's smallest order size bends the ladder. Each rung is about a quarter of the money's worth. On BTC that rounds down to steps of 0.001 BTC, so the ladder held only 87–88% of what it was meant to (72% at worst), and its rungs were uneven by up to 2 times.
What it means
- A wider range is less risk, not shown to earn. What held on both windows is the risk side: fewer stops and smaller worst months. At 30% either side the worst month was smaller than at 15% on 13 of 14 coins without the stop.
- The stop's distance is a choice of how much risk to take. Closer means smaller worst cases and more frequent stops.
- The firing rule isn't worth changing on this evidence. Its gain is one wick.
- Rung count doesn't matter at these gaps.
Limits
- 48 months is few, and three-month stretches overlap.
- The stop fires in a handful of months, so every rule comparison rests on 2–7 events.
- Binance's prices stand in for the exchange the fees come from.
- ETH doubled over the four years, which helps a long grid.
Update, 16 October 2026: a clarification
Added after this page was published. The text above is left as it was.
- Finding 6's settings, named. Its table is 8 rungs with the stop 7.1% under the bottom on the 5-minute rule. The 14-coin counts below it (1 of 14 at 25%, 2 at 30%) are the same setting. With no stop, 4 of 14 were positive in both halves at each of those widths.
- Why finding 2's ETH figures differ from finding 6's. Finding 2 averages over the other settings (rung counts and stops), and finding 6 is one setting.
Update, 21 October 2026: plain words
Added after this page was published. The text above is left as it was, except the explanations this lists.
- Now explained in plain words, with a small example, where each first appears: wicks and mark price.
Credits
- R. A. Fisher, The Design of Experiments, Oliver and Boyd, 1935: the permutation test behind the shuffled months.
- The range widths, the stop rules and the four-year test are our own design.
- Data: Binance's public hourly prices and one-minute mark prices; Bybit's public funding history, fees and order steps.
- PerpGremlin and Claude, Grid bots, followed up: range width, the stop, and four years of ETH, TerminalGremlin, 2026: the tests, the code and the words.
← Will a price touch a level?Grids with and without stops →
Educational research from public market data; any trading discussed is paper, never real. Nothing here is advice. The full notice.