Graveyard / Calendar effects / Binance perps
Crypto tax-loss rebound at year end
Do the year's worst crypto losers rebound in January?
The idea
Investors sell losers late in the year for tax reasons and the selling pressure lifts in January. Buy the coins with the worst year-to-date return on 20 December and hold until mid-January.
How we tested it
| Data | Binance USDT perpetuals, daily closes built from hourly bars, including delisted names. Universe 25 coins in 2020 up to 468 in 2025. Bottom quintile by year-to-date return on 15 December, equal weight. Four windows were fixed in advance, hedged (short BTC perpetual) and unhedged. Primary: 20 December to 15 January, hedged. |
|---|---|
| Period | Six turns of the year, 2020/21 to 2025/26. Development: first four. Holdout: 2024/25 and 2025/26. |
| Costs | 5 bp per side taker on every leg, 7.5 bp as a stress case, plus the funding actually paid or received. |
| Holdout | The last two turns, looked at only after the development table was printed. |
| Control | Calendar placebo (same construction in each of the other eleven months) and a random-quintile twin with 500 draws per season. |
Result
| Hedged mean, all six turns | +0.4% (4 of 6 positive) |
|---|---|
| Hedged mean, development / holdout | +2.2% / −3.2% |
| Standard deviation of one season | 21.7% |
| Calendar-placebo bootstrap p | 0.39 |
| Random-quintile twin p | 0.58 |
| Unhedged mean, all turns | +12.4% (random quintile +14.5%) |
Why it failed
The hedged result is flat. The unhedged gain is crypto beta: random names of the same universe did as well or better. With six turns the test could only detect an effect of about 25% per season, so it can kill a large effect but cannot prove a small one absent.
What would make it worth retesting
Only after at least 10 turns of the year have accumulated, or on a venue with real tax-loss selling in altcoins and a long clean history.
Test id: taxloss-crypto-event-v1. Past results of a backtest say nothing certain about the future.