// research
Crypto momentum after the paper: a weekly test on 8 coins
Partly. Trading on last week’s return beat buy & hold’s Sharpe ratio in only 9 of 56 coin-and-weekday runs. Looking back four weeks did better — a higher Sharpe in 33 of 56 and a smaller drawdown in all 56 — but moving the decision by a day shifted BTC’s Sharpe anywhere from 1.00 to 1.47.
8 min read
In 2018 Yukun Liu and Aleh Tsyvinski showed that cryptocurrency returns are driven by factors of their own rather than by stocks, currencies or precious metals — and that one of the strongest was time-series momentum: a good week for bitcoin tended to be followed by more good weeks.1 Their data ended in May 2018. Everything from 2020 on is therefore a test the authors could not have seen, which makes it a fair question: did the effect survive its publication?
The answer depends on the look-back. The paper's own horizon — this week's return — did not hold up: as a long-or-flat rule it trailed simply holding the coin on a risk-adjusted basis on almost every coin and every weekday. Two to four weeks did better: about buy & hold's Sharpe ratio while in the market half the time, and with smaller drawdowns. But how much better depends on which day of the week the decision is made, by more than the gap to holding itself.
What the paper found
- Data: bitcoin from 2011-01-01, XRP from 2013-08-04 and ether from 2015-08-07, all to 2018-05-31; weeks run from Sunday to Sunday, UTC.
- Momentum: for bitcoin, this week's return predicted the returns 1, 2, 3 and 4 weeks ahead; sorting weeks into fifths, the best fifth of weeks was followed by much stronger weeks than the worst fifth. For XRP the effect showed up over 1–5 days, and for ether it was weaker.
- The paper reports predictability, not a trading rule with costs — so a rule has to be built to check it.
Data and method
- Market: Binance spot, daily candles, 2020-01-01 → 2026-10-01 (the platform's earliest start), BTC, ETH and XRP — the paper's three coins — plus BNB, ADA, DOGE, LTC and LINK.
- The rule: once a week, on the close of the Sunday candle — the paper's week — buy if the close is above the close N days earlier, otherwise stay in cash. The decision fills at the next open, Monday 00:00 UTC. N is 7, 14, 21 or 28 days: the paper's one-week return and the 2–4 week horizons it found predictive.
- Simplification: the paper sorts weeks into fifths; a sort needs the whole distribution, so the rule uses only the sign of the return — up or down.
- Costs: 0.1% fee and 0.02% slippage on every buy and sell, 99% of the account in the position, 10,000 USDT.
- Benchmark: buy & hold run through the same engine with the same costs, so its Sharpe ratio and drawdown are computed exactly like the rule's.
- How much was tried: 4 look-backs × 8 coins on the paper's Sunday week, saved as backtests; the same 4 × 8 again for each of the other six weekdays, as unsaved dry runs (224 configurations in all); and the Sunday week once more over 2023-01-01 → 2026-10-01. Nothing was tuned.
The paper's week: one week back did not beat holding
With the paper's own week, the one-week rule had a higher Sharpe ratio than holding on 3 of 8 coins — ETH, BNB and DOGE — and a smaller maximum drawdown on all 8. The longer look-backs did better: 14 days beat holding's Sharpe on 5 of 8 coins, 21 days on 6, 28 days on 4. The 28-day rule also had the smaller drawdown on all 8 coins and was in the market 45–57% of the time.
Total return is a different matter. Holding half the time, the 28-day rule ended above buy & hold on 4 coins and below it on 4:
Backtest
Buy & hold benchmark (spot 1d)BTCUSDT · 1d · Spot · 2020-01-01 → 2026-10-01
- Return
- +1,066.45%
- Max drawdown
- 76.56%
- Sharpe
- 0.91
- Win rate
- 100.0%
- Trades
- 1
Backtest
Weekly time-series momentum, Sunday close, 28-day lookbackBTCUSDT · 1d · Spot · 2020-01-01 → 2026-10-01
- Return
- +3,373.36%
- Max drawdown
- 33.16%
- Sharpe
- 1.47
- Win rate
- 47.5%
- Trades
- 40
Backtest
Weekly time-series momentum, Sunday close, 7-day lookbackBTCUSDT · 1d · Spot · 2020-01-01 → 2026-10-01
- Return
- +463.89%
- Max drawdown
- 55.22%
- Sharpe
- 0.82
- Win rate
- 33.7%
- Trades
- 92
The weekday decides as much as the rule
The paper's week ends on Sunday, but nothing in the idea of momentum is tied to Sunday. The same rule decided on Monday's close instead gave BTC's 28-day version a Sharpe ratio of 1.00 rather than 1.47 and a drawdown of 64.00% rather than 33.16%:
Backtest
Research: TSMOM weekly (Liu & Tsyvinski)BTCUSDT · 1d · Spot · 2020-01-01 → 2026-10-01
- Return
- +856.14%
- Max drawdown
- 64.00%
- Sharpe
- 1.00
- Win rate
- 41.0%
- Trades
- 39
So every rule was run once for each day of the week. The table shows the range of Sharpe ratios over the seven decision days (unsaved dry runs, same costs):
| Sharpe over 7 decision days, 2020–2026 | Buy & hold | 7 days: lowest to highest | median | 28 days: lowest to highest | median |
|---|---|---|---|---|---|
| BTC | 0.91 | 0.44 to 0.85 | 0.66 | 1.00 to 1.47 | 1.21 |
| ETH | 0.96 | 0.43 to 1.09 | 0.79 | 1.08 to 1.43 | 1.23 |
| XRP | 0.78 | 0.31 to 0.85 | 0.60 | 0.58 to 0.85 | 0.75 |
| BNB | 1.13 | 0.52 to 1.19 | 1.04 | 0.94 to 1.36 | 1.14 |
| ADA | 0.78 | 0.20 to 0.84 | 0.60 | 0.84 to 1.15 | 1.02 |
| DOGE | 0.82 | 0.39 to 1.18 | 0.78 | 0.73 to 0.94 | 0.84 |
| LTC | 0.52 | -0.04 to 0.43 | 0.31 | 0.28 to 0.56 | 0.43 |
| LINK | 0.82 | 0.19 to 0.65 | 0.50 | 0.68 to 0.91 | 0.75 |
Counted over all 56 coin-and-weekday runs:
| Look-back | Sharpe above buy & hold | Max drawdown below buy & hold |
|---|---|---|
| 7 days | 9 of 56 | 44 of 56 |
| 14 days | 32 of 56 | 45 of 56 |
| 21 days | 26 of 56 | 48 of 56 |
| 28 days | 33 of 56 | 56 of 56 |
The one-week signal's median Sharpe was below holding on all 8 coins. The 28-day rule's median was above holding on 5 of 8, by margins from 0.01 (BNB) to 0.30 (BTC) — while the spread between its best and worst weekday was 0.47 on BTC and 0.42 on BNB. The 28-day rule made only 35 to 45 trades in almost seven years on the Sunday week, so a single week moved in or out of the market changes the result; that is the noise the gap to holding has to be measured against.
After 2023
| Sharpe / max drawdown, 2023-01-01 → 2026-10-01, Sunday | Buy & hold | 7 days | 28 days |
|---|---|---|---|
| BTC | 1.16 / 52.91% | 0.81 / 29.53% | 1.33 / 32.17% |
| ETH | 0.65 / 67.40% | 0.62 / 42.83% | 1.13 / 38.08% |
| XRP | 0.85 / 71.95% | 0.56 / 69.81% | 0.52 / 57.35% |
| BNB | 0.85 / 58.11% | 0.58 / 39.79% | 0.83 / 43.41% |
| ADA | 0.40 / 88.13% | -0.02 / 83.53% | 0.11 / 80.12% |
| DOGE | 0.50 / 85.07% | 0.99 / 52.54% | 0.58 / 66.46% |
| LTC | 0.33 / 69.82% | -0.07 / 66.25% | -0.03 / 70.26% |
| LINK | 0.71 / 75.27% | -0.02 / 74.96% | 0.79 / 48.85% |
In the second half of the period the one-week rule beat holding's Sharpe on one coin (DOGE) and had a negative Sharpe on three. The 28-day rule beat it on 4 of 8 and kept a smaller drawdown on 7 of 8 — LTC was the exception.
Conclusion
- The paper's headline signal — this week's return — did not carry over to 2020–2026 as a trading rule: long-or-flat on it beat holding's Sharpe ratio in 9 of 56 coin-and-weekday runs.
- Two to four weeks of momentum acted as a risk filter more than a source of return: about holding's Sharpe ratio while in the market half the time, and a smaller drawdown in every one of the 56 runs of the 28-day rule.
- The day of the week the rule is checked moved its result as much as the gap to buy & hold did; a single backtest of a weekly rule says little until the other six days have been run too.
More research: moving average crossovers on crypto and DCA or lump sum. How to read the cards: backtest metrics explained.
Notes
-
Yukun Liu and Aleh Tsyvinski, "Risks and Returns of Cryptocurrency", NBER Working Paper 24877, August 2018; published in The Review of Financial Studies 34(6), 2021, pp. 2689–2727. Data and horizons as described in sections 2 and 4.1 of the working paper. ↩
Test your own idea
Create a free account, start from a template or from your own rules, and read your first report.
More articles
- Does high funding predict crypto crashes? 8 coins, 2020–2026BIS economists found that high crypto carry predicts crashes. On 8 Binance perpetuals, 2020–2026, crashes came more often — but after bigger gains.
- Does price move the hour before funding? 8 coins, 2020–2026Do traders sell before paying funding? The hour before each settlement on 8 Binance perpetuals, 2020–2026: no reliable move, and fees took what was left.
- Bitcoin after the halving: 2020 and 2024 backtestedDid buying BTC at the halving pay? A year after 2020: +554.04%, top 2.5% of start days. After 2024: +30.45%, below the median. Two cases, no pattern.
- Weekends and time of day: when does bitcoin move?Do bitcoin gains come on weekends or outside US trading hours? Hourly BTC and ETH data since 2020, and session backtests: with fees, -99.87%.
- What leverage does to one strategy: 1x, 2x, 3x, 5xThe same golden cross on BTC and ETH perpetuals, 2021–2026: BTC +9.78% at 1x, -7.76% at 2x, -99.77% at 5x. Funding times leverage moved the liquidation price.