Archive / Equities & volatility / E380
Leveraged-fund rebalancing leaves no detectable trace before the close
Trading with the day's move from 15:30 to 16:00 earned +1.9 bp before costs on Bitcoin and Ether, no more than +2.9 bp before leveraged Bitcoin funds existed, against a 14 bp cost.
Leveraged funds that reset daily must buy after a rise and sell after a fall, and they do it at the close. The belief is that this pushes prices further the same way during the last half hour. On days with a move of 3% or more since the previous close, we traded with it from 15:30 to 16:00 New York time. On Bitcoin and Ether the average was +1.9 bp (1 bp = 0.01%) before costs against a 14 bp cost. Before the first leveraged Bitcoin fund the same rule gave +2.9 bp, and stocks with large leveraged funds gave -1.7 bp. None of 81 pre-registered tests passed.
Daily-reset leveraged funds must trade in the direction of the day's move into the close, so price keeps going until 16:00.
Why people believe it
A fund that targets 2x or 3x the daily return resets its exposure to the previous close. After a day with return r, it has to trade about its assets times L(L-1) times r in the same direction as r, and it has to do so by the close, because the next day's target is measured from the closing price. The size is known in advance from public fund assets and the day's return. For the names with the largest funds relative to their daily trading, the flow can be a visible share of the closing half hour.
Why it would persist The argument, as stated before testing
The fund cannot choose when to trade or in which direction, so the flow is a cost of the product rather than a mistake that a trader can compete away. Anyone who wants to front-run it has to hold a position through a volatile half hour and pay a round-trip cost.
For each trading day we measured how far the price had moved since the previous 16:00 New York close. If the move was 3% or more up or down at 15:30, we bought (or sold) in that direction at 15:30 and closed at 16:00, after a 14 bp round-trip cost (1 bp = 0.01%). We did this on Bitcoin and Ether perpetual contracts, which trade around the clock, and on perpetual contracts on single US stocks. To tell a leveraged-fund effect from ordinary behaviour around the close we used controls: Bitcoin and Ether before the first US leveraged Bitcoin fund, and stocks that have no large leveraged funds. We also tried shifted windows, a midday placebo and a next-morning reversal. The test was fixed before any conditional return was computed.
Net mean above +5 bp (gross above 19 bp), date-clustered t of the gross mean at least 2.5, and the control gross mean below half the treatment gross mean.
Net mean at or below +5 bp, or t below 2.5, or the control gross mean at least half the treatment gross mean; applied to the crypto cell and the leveraged-fund stock cell.
14 bp round trip (7 bp per side); stress case 20 bp per side.
Crypto: the same rule on 2022-01 to 2023-06 (before the first leveraged Bitcoin fund) and on spot 2019-09 to 2021-12. Stocks: names without large leveraged funds, plus the exploratory group. Returns are averaged per date and date-clustered bootstrap intervals are used; all 81 cells (9 rules by 11 sets) form one Benjamini-Hochberg family at 5%.
Net return was -12.1 bp (the +5 bp rejection line), the t-statistic was 0.70 (needed 2.5), and the pre-leveraged-fund control showed +2.9 bp, more than half of the +1.9 bp treatment; the stock cell failed the first two checks.
- Bitcoin and Ether, 15:30-16:00, days with a move of 3% or more since the previous close (298 dates): +1.9 bp before costs, t 0.70, 95% interval -3.6 to +7.6 bp, net -12.1 bp. Bitcoin alone +1.2 bp (t 0.36), Ether alone -1.3 bp (t -0.37).
- The same rule from 2022-01 to 2023-06, before the first US leveraged Bitcoin fund (178 dates): +2.9 bp (t 0.47, interval -9.7 to +15.2 bp), larger than the treatment period; the difference between the two periods has t -0.15. On spot data for 2019-09 to 2021-12 (325 dates) it was +0.9 bp.
- Stocks with large leveraged funds (161 dates): -1.7 bp (t -0.35, interval -11.8 to +8.0 bp, net -15.7 bp). Stocks without them (113 dates): -2.6 bp (t -0.65). The difference has t +0.13.
- None of the 81 pre-registered cells passed Benjamini-Hochberg at 5%. The 15:45-16:00 window gave +1.3 bp on crypto and +3.2 bp (t 0.96, 162 dates) on stocks with large leveraged funds; the 11:30-12:00 placebo gave +3.1 bp on crypto and -2.3 bp on those stocks.
- Holding the position overnight in the other direction (16:00 to next 09:30) gave -0.7 bp (t -0.04) on crypto in 2024-2026 and +12.6 bp (t 0.73) on stocks with large leveraged funds, so no reversal leg appeared either.
- The size of the flow was not measured: no fund-assets history was collected, so the 81 pre-registered cells test only the price pattern, not the amount of fund trading.
- Post hoc, crypto: the 15:30-16:15 window lost 9.1 bp before costs (t -2.55), from a reversal after 16:00 of -10.8 bp (t -5.0, 310 dates) in 2024-2026 that was already present in 2022-2023 at -11.2 bp (t -2.9). This window was looked at after the test, so it is not evidence.
- Post hoc, stocks (about 90 post hoc cells, no multiple-testing correction): in the five minutes after 16:00, stocks with large leveraged funds moved +3.1 bp (t 2.0) in the direction of the day, against -5.1 bp (t -3.5) for stocks without them; most of the group difference is the comparison stocks giving back 5.1 bp. The minimum detectable effect for the stock cell was 16.1 bp, so effects of a few bp could not be ruled out. These cells were looked at after the test, so they are not evidence.
The tested window shows nothing, with or without leveraged funds
Mean return of trading with the day's move from 15:30 to 16:00, per date, with date-clustered 95% intervals. Every interval includes zero, the crypto control period is not below the treatment period, and the pass line is +19 bp before costs.
Chart data (bp, before costs)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Bitcoin and Ether 2024-2026 (298 dates) | 1.9 | -3.6 | 7.6 |
| Bitcoin and Ether 2022-23, before leveraged funds (178 dates) | 2.9 | -9.7 | 15.2 |
| Stocks with large leveraged funds (161 dates) | -1.7 | -11.8 | 8 |
| Stocks without large leveraged funds (113 dates) | -2.6 | -10.4 | 4.6 |
Return by window around the close, post hoc
Stocks with large leveraged fundsStocks without large leveraged funds
Return in the direction of the day's move by window, on days of 3% or more. Neither group moves in the last 30 minutes; the groups differ only after 16:00 (+3.1 bp against -5.1 bp), mostly because the comparison stocks give back 5.1 bp (t -3.5). The 15:45-16:00 value (+1.9 bp) differs from +3.2 bp (162 dates) in the cell table because the two analyses use different samples. Chosen after seeing the results.
Chart data (bp, before costs)
| Stocks with large leveraged funds | Stocks without large leveraged funds | |
|---|---|---|
| 15:30-15:45 | -3.7 | -1.2 |
| 15:45-16:00 | 1.9 | -1.3 |
| 16:00-16:05 | 3.1 | -5.1 |
Why it failed Interpretation · not established
The data show no continuation in the continuous session before the close, in crypto or in stocks. The post hoc pattern puts whatever footprint exists at the 16:00 closing auction and the minutes after it, not in the half hour before. Names without large leveraged funds give back about 5 to 12 bp after 16:00, while names with them do not, which suggests that the flow is executed in the auction rather than spread over the last half hour. This is an interpretation: the test did not observe auction orders. In crypto, the 16:00 reversal also appears before any leveraged Bitcoin fund existed, so it is generic and not a fund effect.
Lesson
A mandatory flow whose size can be computed in advance is not the same as a price move in a particular time window. Where the flow is executed matters as much as how large it is, and a control group without the flow is needed to separate it from a reversal after 16:00 in the comparison stocks, which have no large leveraged funds.
Scope of this result
Covers Bitcoin and Ether perpetuals (2022-2026), nine single-stock perpetuals with large leveraged funds against 35 without, and the two index perpetuals, 30-minute and nearby windows, a 3% threshold (1.5% for the index perps), and a 14 bp round-trip cost. It does not cover the underlying stocks or the leveraged funds themselves, closing-auction orders, other thresholds chosen in advance, or fund assets, which were not measured. The post hoc cells are descriptive only.
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-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 5Finding 6Finding 7Finding 83. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade