The Lab

How we tested

• Bitcoin, Ether and Solana futures, October 2022 to September 2026, checked every minute.
• The bot as claimed: always in a trade; the moment one closes, the next opens. Run buying only, selling short only, and picking the side by coin flip.
• Real costs: the exchange's market-order fee (0.055%) to get in and out, and the funding that futures pay or charge while a trade is open.
• Cautious fills:
  • if the price jumps past the stop, the stop fills at the worse price;
  • if one minute touches both the stop and the target, we count it as the stop.
• Three comparisons:
  • random entries: wait a random time, then trade a random side;
  • shuffled prices: each month's hours in random order (a permutation test, an idea from the statistician R. A. Fisher);
  • driftless prices: each month tilted so it ends where it began, which removes the market's trend and keeps its swings.
• The tuning, done fairly:
  • 720 versions per coin: stops from 0.5% to 5%, targets from half to three times the stop, buying, selling or both, with or without a simple trend filter or a 24-hour time limit;
  • tuned on the first three years only;
  • the winner then judged once on the last year, which the tuning never saw.

The maths, step by step

1. The odds of a 2/4 bracket. With the price at 100, the stop is at 98 and the target at 104. On a price with no trend, the chance of reaching the target first is the distance to the stop divided by both distances (in percentage terms): 2.02 ÷ (3.92 + 2.02) = 34.0%, about 1 in 3. This is the old "gambler's ruin" result, worked out by Huygens and others in the 1600s, not ours.

2. The average trade before fees. 34.0% × 4% − 66.0% × 2% = +0.04%. That is nearly nothing.

3. After fees. Getting in and out costs 0.055% twice, 0.11%. So +0.04% − 0.11% = −0.07% a trade.

4. What it would need. To break even after fees, the target has to come first (2% + 0.11%) ÷ 6% = 35.2% of the time.

What we found

1. The bot hits its target about 1 time in 3, just as the maths says.

• On the driftless prices, every coin and side hit 32.8% to 34.4%, and every one lost.
• Random entries on random sides hit 33.4%.

2. Where it made money, the market's rise made it.

Each figure is the sum of every trade's result on one full-size position, so −100% means a whole account.

• Only Bitcoin, buying, ended up: it hit its target 35.9% of the time, just past break-even, in four years when Bitcoin more than quadrupled.
• In the last year, when the coins fell 27% to 43%, buying lost on every coin (−56% to −97%).
• Shuffling the order of the hours changed nothing: the order of moves isn't what the bot trades.

3. Tuning found the trend, not an edge.

• Nine of the ten best versions on the first three years bought with wide stops (3% to 5%) and no filter. Those versions trade least, so pay the least fees, and spend the most time riding the rise.
• The winner (stop 5%, target 15%, buying) made +149% to +348% in those years. In the year it hadn't seen, it lost on all three coins: −27% Bitcoin, −26% Ether, −48% Solana.
• On the driftless version of that year it made +3%, +12% and −0.3%: no edge once the trend is taken out.
• A standard test for picking the best of many tries (the deflated Sharpe ratio, from Bailey and López de Prado, 2014) gave it a 0% chance of being more than luck.

4. What each setting did. Tighter stops lost more, because more trades means more fees. Wider targets and the trend filter helped a little; neither turned it into a profit.

Credits

• Christiaan Huygens, De Ratiociniis in Ludo Aleae (On Reasoning in Games of Chance), 1657: the gambler's ruin problem behind step 1.
• R. A. Fisher, The Design of Experiments, Oliver and Boyd, 1935: the permutation test behind the shuffled prices.
• David H. Bailey and Marcos López de Prado, "The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality", Journal of Portfolio Management, 2014.
  • Overfitting is tuning a rule until it fits the past's noise rather than a real pattern, so it fails on new prices. For example, of 100 settings with no skill, the best on last year's prices will usually look good by luck, and do no better than the rest the year after.
• The claim tested: a quant's AI-run strategy sweep, as it was passed round. We don't know its original author, so we don't name one.
• The driftless and random-entry controls, and the tuning test, are our own design.
• Data: Binance's public 1-minute futures prices; Bybit's public funding history.
• PerpGremlin and Claude, Does a "2% stop, 4% target" bot beat every strategy?, TerminalGremlin, 2026: the tests, the code and the words.