Archive / Crypto & DeFi / E821
Shorting a token as large stakes end shows no detectable edge
Across 1,132 large scheduled stake endings, a 24-hour hedged short after the end lost 4.6 bp net and one before it earned +6.9 bp; both intervals include zero and the largest endings did worst.
On a proof-of-stake network with fixed stake end dates, the hour at which staked tokens return to their owners is public well in advance, and many traders expect some to be sold. We found 1,132 clusters in which at least 250,000 tokens were due back within one hour, and shorted the token's perpetual future for 24 hours, hedged with Bitcoin. After 10 bp of costs the short started 15 minutes after the end lost 4.6 bp on average (90% interval -20.9 to +11.7). Shorting the 24 hours before the end earned +6.9 bp, also indistinguishable from zero. The largest third of endings did worst, at -42.2 bp.
When a large amount of staked tokens reaches its scheduled end, the owners receive liquid tokens and some sell them, so the price falls around the end and a short earns a premium.
Why people believe it
A stake with a fixed term returns the principal and the reward at an end time that has been known since the stake began. People who staked only to earn the reward, or who wanted out, can sell at once, and the tokens arriving in an hour can be large relative to normal trading.
Why it would persist The argument, as stated before testing
The end times are public, but the amount ending in any given hour is lumpy, and derivatives market makers may not position for it in advance.
We downloaded every completed validator stake on the network from 2022 to 2026 and summed, hour by hour, how many tokens were due back. An hour with at least 250,000 tokens ending counted as an event, and events within 24 hours of an earlier one were merged into the first. We then simulated a 24-hour short of the token's perpetual future, hedged with an equal-size long in Bitcoin so that general market moves cancel. The short started either just after the stakes ended or 24 hours before. A placebo repeated the same short at the same clock time one week earlier.
H1 passes only if all hold: at least 30 clusters; hedged net mean above 0 with the lower end of a bootstrap 90% interval above 0; median hedged gross above 0; mean of the largest dose third above that of the smallest third; placebo mean below half the event mean. H2 uses the same gates. Two hypotheses, no further correction.
Any gate fails.
10 bp round trip. Funding is not modelled.
Placebo: the same trade at the same clock time 7 days earlier, kept only when no event lies within 24 hours of that time (110 of 1,132 clusters). Raw short without the Bitcoin hedge reported alongside.
The mean itself is negative (-4.6 bp). The main test fails three pre-registered gates: the 90% interval's lower end is -20.9 bp, the largest third (-42.2 bp) does worse than the smallest (+7.6 bp), and the placebo (+29.5 bp) is not below half the event mean. The anticipation test fails the interval and placebo gates.
- Shorting 15 minutes after the end, hedged with Bitcoin, lost 4.6 bp net on average over 24 hours (n 1,132; 90% bootstrap interval -20.9 to +11.7 bp). Without the Bitcoin hedge the net mean was -9.1 bp.
- The median trade made +21.7 bp gross (+11.7 bp net) and 53.1% of trades were profitable net, but the worst single trade lost 2,636 bp net, which pulls the mean below zero: the typical trade is not the average, and the worst trade outweighs the typical gain. The median and the win rate are not pass criteria.
- Anticipation: shorting from 24 hours before the end to the end earned +6.9 bp net (n 1,131; 90% interval -9.6 to +23.0 bp), with the interval's lower end below zero.
- By dose, the short after the end earned +7.6, +19.4 and -42.2 bp in the smallest, middle and largest thirds (about 375 clusters each); the pre-registered rule needed the largest third to beat the smallest. For the anticipation short the thirds were +2.4, -0.9 and +18.2 bp: non-monotonic and within noise (standard error about 17 bp each). One incidental gate pass across ten gates is not evidence.
- The placebo, the same short at the same clock time one week earlier, earned +29.5 bp (n 110; 90% interval -24.7 to +81.2 bp), more than either event test.
- By year the short after the end earned +16.0 bp (2022, n 199), -48.6 (2023, n 166), +1.7 (2024, n 243), +1.5 (2025, n 268) and -4.5 bp (2026, n 256); only the 2023 interval lies entirely below zero.
- Re-staking: the median ending cluster had 82% of its tokens staked again by the same node within 48 hours. The half of clusters with a lower re-stake share lost 5.2 bp and the other half lost 4.0 bp (566 clusters each).
No edge before or after the end
Mean 24-hour net return of the Bitcoin-hedged short with its 90% bootstrap interval. Both event tests straddle zero; the placebo sits higher than either, with a very wide interval because only 110 placebo times were free of events.
Chart data (bp)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Short after the end (n 1,132) | -4.6 | -20.9 | 11.7 |
| Short before the end (n 1,131) | 6.9 | -9.6 | 23 |
| Placebo: same time, 7 days earlier (n 110) | 29.5 | -24.7 | 81.2 |
No positive dose response
Net return of the short after the end, by size of the ending relative to daily perpetual volume (thirds of 375 clusters). The largest third is lower than the others, but within noise; the 2023 result alone is -48.6 bp.
Chart data (bp)
| Short after the end | |
|---|---|
| Smallest third (up to 0.12 of daily volume) | 7.6 |
| Middle third (0.12-0.33) | 19.4 |
| Largest third (above 0.33) | -42.2 |
No stable sign across years
Mean net return of the short after the end by calendar year with 90% bootstrap intervals. The sign changes from year to year and every interval except 2023 includes zero.
Chart data (bp)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| 2022 (n 199) | 16 | -28.2 | 60.5 |
| 2023 (n 166) | -48.6 | -98.9 | -0.5 |
| 2024 (n 243) | 1.7 | -35.2 | 37.6 |
| 2025 (n 268) | 1.5 | -32.3 | 35.5 |
| 2026 (n 256) | -4.5 | -27.7 | 18.3 |
Why it failed Interpretation · not established
The event study found no detectable selling pressure in price at this power (standard error about 10 bp, validator stakes only). Most large endings are stakes that the same nodes start again (median 82%), and the clusters with less re-staking show the same result. The largest third of endings is lower than the others, but within noise. One possible reading is that holders of an already owned, already hedgeable token have no deadline to sell.
Lesson
A known date on which a token becomes liquid is not the same as a date on which it must be sold. Supply events move prices only when the recipient faces a deadline or a cost of holding; when the holder can wait, the timing of the unlock says little about the timing of the sale. A related test of scheduled token unlocks (E390) likewise found no detectable drift.
Scope of this result
Covers validator stakes on one network from January 2022 to October 2026, endings of at least 250,000 tokens per hour, a 24-hour hold with a Bitcoin hedge and 10 bp round-trip costs, on hourly buckets and 15-minute prices. Delegated stakes are not included, so endings by delegators are not tested.
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 73. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade