The Lab

How we tested

• The book: BTC, ETH and SOL at equal weights, held as coins, each with its own coin-settled short, checked daily and resized past a 5% drift. Funding buys more coins. Bybit's public hourly prices and funding, 1 November 2024 to 1 November 2026.
• 120 rules on each book. Take profit at 2, 3, 5, 8, 12 or 20%; closing all or half of the short; 10 ways to short again: back at the old entry, with no stop or with a stop at a further fall of 10 or 20%; a bounce of 2, 5 or 10%; after 24, 72 or 168 hours; or funding turning positive.
• Two books: no loan, and 3x with a USDT loan at an assumed 5% a year (past loan rates are not in the data). The loan book can be liquidated.
• Costs: 0.055% on each take profit and re-short, 0.02% on routine resizes, 0.1% on the coins.
• Fixed before the first run: one rule was picked per book on the first year alone, then judged on the second, with and without its last 14 days, because that fortnight fell and falls flatter shorts.
• To pass, on both versions of the second year, a rule must beat neutral on return, match it on return over worst drawdown, and beat 97.5% of shuffled timings of its own spells. It must also never be liquidated, and clear the reality check over all 120 rules (p under 0.05).

What we found

No rule passed the bar, on either book. The rule each book picked did well in the second year, then failed badly on older prices.

Both books picked the same rule: take profit at 20%, close it all, short again after 24 hours. Any one figure for it depends heavily on the path, so we start with the spread. We ran the same books from each of 24 start hours, so that only the hour of the daily check changed. These runs share the same few events, so they are not 24 samples.

• So the second year's result is not only one lucky start. In that year, buying after a 20% fall and hedging again a day later usually gained.
• The older years undo it. The choice never saw these years, and they are reported, not judged. With the loan the rule was liquidated from all 24 starts: -69.06% from the first start, against neutral's -11.81%. With no loan it made -8.28% with a 33.67% drawdown, against neutral's +10.37% and 14.89%. SOL's funding there is partly its USDT market's, for the stretch when no coin-settled market existed.
• Liquidations: 18 of the 120 loan rules were liquidated within the two years, all of them closing the whole short. Those rules can sit long at 3x through a fall.

The bar, judged as registered from a single start (the first hour of the data). It took 4 take profits in the year.

The 3x rule's figure in that table has read three different values through small corrections to our model (see Limits). Read it as one draw from the spread above.

• Luck explains the best of them. The reality check over all 120 rules gives p = 0.62 with no loan and 0.78 with the loan. The bar was under 0.05.
• The first year did not predict the second. Across the 120 rules, the rank correlation between the two years was -0.13 with no loan and +0.05 with the loan.
• No loan: from this start the rule made a little less than neutral, with about four times the drawdown. The first year looked very different: +25.93% against +7.60%.
• 3x loan: from this start the rule cleared the tests on return and on return over drawdown. That bar is low here. Neutral lost money because the 5% loan cost more than funding paid, and 45 of the rules cleared both tests on both versions of the year. The rule beat all 1,000 random placements of its own spells, but only 95.9% of shuffled timings, under the 97.5% bar. With so few take profits that control has just 48 arrangements, and the rule's own ranked second of them.

What kind of market it needs. Across all 120 rules with no loan, compared with neutral month by month:

That is the one-rung grid's shape. It wins when falls come back and loses when they do not.

What it costs, median of the 120 over two years with no loan: unhedged 4.8% of the time, fees of 1.28% against neutral's 0.11%. Funding given up came out slightly negative, at -0.37%. These books often ended up larger, and a larger book earns more funding, so that does not mean being unhedged was free.

A correction this study found

The first run showed a rule collecting more funding than neutral while almost always hedged. The cause was in our models of the basket: a coin-settled short's profit, loss and funding are paid in the coin, and they were booked in dollars. A short closed at a loss became a fixed debt, at no interest, which let the book grow against its money. One model also took the coins' 0.1% buying fee twice. All of this is fixed, and this page uses the fixed count.

The neutral basket over this page's two years goes from ×1.107 to ×1.101. We then measured which earlier figures move:

• Unchanged: the two-year ×1.13 on "Funding carry and the hedged basket" (×1.136 to ×1.133, re-run on today's engine). "The same long, held two ways" is unchanged too: its books never resize and keep funding as cash, so all 50 rows match.
• Moved: the loan, the funding rule, the tilts and the sideways table, all on the carry page. Its dated update of 3 November 2026 shows them beside the published figures. What it concludes still holds.

The verdict here was also no on the model as first registered.

Limits

• Results here swing with the path. While checking this study, two small changes moved the 3x rule's second year from +10.10% to -0.99% to +12.16%, with nothing changed in the rule: matching the basket model's rebalancing test, then paying funding in the coin. Read the spread above, not any one figure.
• Neutral itself moves by about a point over two years with which rebalance happens to fire. Differences of that size are noise.
• Hourly closes only: a level touched and left within an hour does not fire.
• The coins are priced by the USDT perpetual, so the gap between them and the coin-settled short is not modelled.
• The loan rate is an assumption, and the loan's margin terms partly so.
• The fixed count, random placement and the 24 starts were added after seeing the first run, and are reported only. The way a rule is picked and the bar did not change.
• 480 variants were run in all: 120 per book, on each of the two models.
• These prices are not new to us, as earlier studies walked much of them.