Be in the coin while it trends. Sit in cash while it doesn't.
The rule holds BTC and ETH only while each one closes above its own 50-day average, and sizes the position so a wild market gets a smaller one. It is old, simple and public - and out of 185 strategies we tested, it was the one that beat holding Bitcoin on risk-adjusted terms in years it had never seen.
I.How it works
Three decisions a day, the same for everyone.
- Is the trend up? After the 00:00 UTC close, the coin's close is compared with the average of its last 50 closes. Above: risk-on. Not above: risk-off, weight 0 - that share of the book is cash until the close is back above the line.
- How big? Risk-on size = the smaller of 100 % and 60 % divided by the coin's annualised volatility over the last 30 days. A calm uptrend gets a full share; a violent one (say 120 % volatility) gets half.
- What does it cost? The weight is held for the next day. Every change pays 0.15 % per side of the amount moved (fee plus slippage) in the test; the record charges the same.
Trends persist
Time-series momentum - assets that rose tend to keep rising for a while - is one of the best documented effects in markets (Moskowitz, Ooi and Pedersen, 2012), and it has been measured in crypto too (Liu and Tsyvinski, 2021). The rule does not predict; it only stays with a move that is already happening.
Volatility sets the size
Sizing by volatility (Moreira and Muir, 2017) keeps a single blow-up week from dominating the year. Crypto's volatility swings between 40 % and 150 % a year; a fixed-size position takes those swings in full.
Cash is a position
The rule's whole edge over holding came from what it skipped: most of 2018 and 2022. In return it is late into every rally and trails holding in the strongest bull years - both are on the charts.
II.What was tested
185 strategies, one rulebook, fixed before the first run.
The protocol (research/lab/PROTOCOL.md, committed before any result existed) fixed the families, their parameter grids, the costs, the periods and the selection rule. Every cell of every grid counts as a trial.
| Family | Idea | Trials |
|---|---|---|
| Trend + vol cap (the v1 rule) | Close above its 50-day average, 60 % volatility cap, BTC+ETH | 1 |
| Time-series momentum | Every coin of the point-in-time top 10 or 30 with a positive 20-180-day return, per-coin volatility targeting, with and without a BTC-trend gate | 96 |
| Cross-sectional momentum | Rotate weekly into the 3, 5 or 10 strongest coins of the top 30, inverse-volatility weights, BTC-trend gate on/off | 24 |
| Breakout + ATR trail | Donchian channel breakout (20-100 bars), exit on a 2-5 ATR trailing stop; BTC/ETH/SOL on daily and 4-hour bars, and the top 10 daily | 48 |
| Regime gates | BTC above its 100-day average; half size while BTC volatility is in its top 20 % of the year | 1 |
| Funding | Skip coins with overheated perpetual funding; a funding-carry sleeve (spot long, perpetual short); idle cash in that sleeve | 7 |
| Machine learning | LightGBM on 30 features (returns, volatility, trend distance, volume, BTC state, funding, open interest, cross-sectional ranks), retrained every January, holds the top 3/5/10 predictions weekly | 6 |
| Capped leverage | The selected rule at 1.5x and 2x gross, excess carried in perpetuals at their real funding | 2 |
Data: daily candles for all 668 USDT pairs Binance ever listed, including the ones later delisted (LUNA, FTT and hundreds more); 4-hour candles; Binance perpetual funding since 2019; Bybit open interest since 2020. The top 30 on any day is ranked from that day's trailing traded value only, so the baskets hold what was big then. A delisted coin is sold at its last close.
III.How the winner was chosen
Picked on 2018-2022. Judged on 2023 to today.
Each family's parameters were chosen on 2018-2022 by the best neighbourhood Sharpe ratio - the cell and its grid neighbours together, so a lucky spike does not win. The family winners were then ranked by that in-sample Sharpe, and the first one down the ranking that beat holding BTC out of sample on Sharpe, Sortino and MAR, with a smaller worst drawdown, at normal and at double costs, and that can be computed every day from the forward record's data (Kraken daily candles), was selected.
That took 8 of the 9 finalists. The funding-carry sleeve passed the numbers - a steady few percent a year with tiny drawdowns in the test - but it needs perpetual-futures funding data and is not a trend rule; the 4-hour breakout came within a hair of BTC's Sharpe but needs intraday bars. Both are reported, neither is selected.
| Family winner (chosen on 2018-2022 only) | IS Sharpe | OOS per year | OOS Sharpe | OOS Sortino | OOS worst DD | OOS MAR | OOS Sharpe 2x costs | Verdict |
|---|---|---|---|---|---|---|---|---|
| Funding carry (spot long + perp short) carry BTC+ETH enter > 0 %/yr | 5.76 | +3.0 % | 8.30 | 14.02 | -1 % | 2.72 | 4.31 | fails: needs data beyond daily candles |
| Donchian breakout + ATR trailing stop ('BTC/ETH/SOL 4h', 100, 4) | 1.71 | +26.6 % | 1.15 | 1.96 | -27 % | 0.98 | 0.98 | fails: Sharpe, MAR, 2x costs, needs data beyond daily candles |
| Cross-sectional momentum (strongest K of the top 30) (14, 5, 'on') | 1.49 | +70.8 % | 1.09 | 1.81 | -67 % | 1.06 | 0.93 | fails: Sharpe, Sortino, deeper drawdown, 2x costs |
| F1 + idle cash in funding carry F1 winner + idle cash in carry (enter > 0 %/yr) | 1.30 | +20.0 % | 1.09 | 1.72 | -23 % | 0.85 | 0.87 | fails: Sharpe, Sortino, MAR, 2x costs, needs data beyond daily candles |
| LightGBM, 30 features, walk-forward (5, 'on') | 1.27 | +32.7 % | 0.89 | 1.33 | -48 % | 0.69 | 0.71 | fails: Sharpe, Sortino, MAR, 2x costs, needs data beyond daily candles |
| F1 + funding filter F1 winner, skip funding > 100 %/yr | 1.15 | +17.8 % | 0.99 | 1.55 | -24 % | 0.73 | 0.77 | fails: Sharpe, Sortino, MAR, 2x costs, needs data beyond daily candles |
| Time-series momentum, top-N basket, vol targeted (20, 0.4, 10, 'on') | 1.10 | +17.9 % | 0.99 | 1.56 | -24 % | 0.74 | 0.78 | fails: Sharpe, Sortino, MAR, 2x costs |
| v1 rule: close above 50-day average, 60 % vol cap (BTC+ETH) sma-50-volcap BTC+ETH | 1.10 | +52.8 % | 1.41 | 2.34 | -30 % | 1.75 | 1.32 | selected |
| F1 + BTC volatility gate F1 winner + BTC vol gate | 1.05 | +17.1 % | 0.99 | 1.57 | -24 % | 0.71 | 0.77 | fails: Sharpe, Sortino, MAR, 2x costs |
| BTC, bought and held (the bar) | 0.44 | +54.5 % | 1.17 | 1.83 | -53 % | 1.03 | - | - |
IS = in sample (2018-2022; the ML model's own walk-forward predictions 2020-2022). OOS = 2023-01-01 to 2026-09-25. Every family ran with 0.10 % fees plus 0.05 % (BTC, ETH) or 0.10 % (other coins) slippage per side.
IV.Results hypothetical
Same pace as BTC since 2023, a much smaller fall.
| Out of sample, 2023-01-01 to 2026-09-25 | Per year | Sharpe | Sortino | Worst drawdown | MAR |
|---|---|---|---|---|---|
| Selected rule (spot, BTC+ETH) | +52.8 % | 1.41 | 2.34 | -30 % | 1.75 |
| BTC, bought and held | +54.5 % | 1.17 | 1.83 | -53 % | 1.03 |
| BTC+ETH, bought and held | +40.4 % | 0.91 | 1.38 | -60 % | 0.67 |
| Top-30 coins, held (point in time) | -21.2 % | 0.03 | 0.04 | -90 % | -0.24 |
| Same rule at 1.5x gross cap (research variant) | +75.2 % | 1.35 | 2.22 | -44 % | 1.73 |
| Whole test, 2018-01-01 to 2026-09-25 | Per year | Sharpe | Sortino | Worst drawdown | MAR |
|---|---|---|---|---|---|
| Selected rule (spot, BTC+ETH) | +46.9 % | 1.22 | 1.90 | -53 % | 0.88 |
| BTC, bought and held | +23.4 % | 0.65 | 0.94 | -81 % | 0.29 |
| BTC+ETH, bought and held | +22.4 % | 0.65 | 0.92 | -88 % | 0.25 |
| Same rule at 1.5x gross cap (research variant) | +64.1 % | 1.17 | 1.81 | -70 % | 0.92 |
Growth of 1, rule vs BTC held
log scaleruleBTC held
Drawdown from the last high
rule vs BTC heldruleBTC held
Calendar years
| Year | Selected rule | BTC held | Difference |
|---|---|---|---|
| 2018 | -22 % | -72 % | +51 % |
| 2019 | +64 % | +94 % | -30 % |
| 2020 | +267 % | +302 % | -35 % |
| 2021 | +105 % | +60 % | +45 % |
| 2022 | -39 % | -64 % | +25 % |
| 2023 OOS | +91 % | +156 % | -65 % |
| 2024 OOS | +50 % | +121 % | -72 % |
| 2025 OOS | +35 % | -6 % | +42 % |
| 2026 (to 09-25) OOS | +26 % | -4 % | +30 % |
V.Robustness
Would it survive a different draw of history?
Monte Carlo on the selected rule
| 10,000 block-bootstrap paths, 2023+ | Per year 5 / 50 / 95 % | Sharpe 5 / 50 / 95 % | Worst drawdown 5 / 50 / 95 % | Paths losing money |
|---|---|---|---|---|
| Selected rule | +11 % / +48 % / +102 % | 0.47 / 1.31 / 2.17 | -46 % / -29 % / -20 % | 1.2 % |
| BTC held | +1 % / +49 % / +123 % | 0.24 / 1.09 / 1.97 | -68 % / -45 % / -31 % | 4.7 % |
10,000 paths re-drawn from 2023+ daily returns in 20-day blocks (keeps volatility clusters together). Reshuffling the order of the rule's 169 trades gives a worst drawdown between -18 % and -37 % in 90 % of orders.
Parameter grids: how much of each family beat BTC
| Family grid | Cells | Beat BTC Sharpe in sample | Beat BTC Sharpe out of sample | Beat BTC MAR out of sample | Median OOS Sharpe |
|---|---|---|---|---|---|
| Time-series momentum, top-N basket, vol targeted | 96 | 88 % | 0 % | 1 % | 0.54 |
| Cross-sectional momentum (strongest K of the top 30) | 24 | 71 % | 0 % | 4 % | 0.30 |
| Donchian breakout + ATR trailing stop | 48 | 100 % | 15 % | 21 % | 1.00 |
| LightGBM, 30 features, walk-forward | 6 | 100 % | 0 % | 0 % | 0.80 |
Most momentum, rotation and ML variants beat BTC in 2018-2022 and none of them did from 2023: altcoins as a group fell far behind Bitcoin in this cycle (the top-30 basket, held, went -21 % a year while BTC grew 5.1-fold), and any strategy that spread across them paid for it. That is the main reason the BTC+ETH rule won.
Walk-forward: parameters re-chosen every January
| Walk-forward (re-chosen every 1 January) | OOS per year | OOS Sharpe | OOS worst DD | OOS MAR |
|---|---|---|---|---|
| Time-series momentum, top-N basket, vol targeted | +13.0 % | 0.79 | -24 % | 0.54 |
| Cross-sectional momentum (strongest K of the top 30) | +70.8 % | 1.09 | -67 % | 1.06 |
| LightGBM, 30 features, walk-forward | +32.7 % | 0.89 | -48 % | 0.69 |
| Donchian breakout + ATR trailing stop | +25.4 % | 1.09 | -27 % | 0.94 |
Luck after 185 tries
The deflated Sharpe ratio (Bailey and López de Prado) asks how likely a Sharpe ratio is to be real after this many trials. For the selected rule it is 0.31 on 2023+ and 0.09 on 2018+ (1.0 = certainly real). In plain words: the backtest does not prove an edge. The rule was already published and running before this test, which is some protection, and the sealed forward record is how it will be proven or disproven.
VI.The 1.5x variant
More return from the same entries - with more risk.
Allowing the same book up to 1.5 times its size (the excess held in perpetual futures and charged their real historical funding) lifted the 2023+ yearly growth above holding BTC while its worst drawdown stayed smaller than BTC's; at 2x the drawdown was deeper than BTC's and it failed the bar. Leverage magnifies losses as well as gains, adds liquidation risk and funding costs that can turn against a holder. It is shown as research only; the published record and every chart run the spot rule.
VII.Limits
What this research cannot tell you.
- 2023 to today is one market cycle. A long sideways market with sharp reversals is where a 50-day trend rule loses most; the out-of-sample years had only short stretches of it.
- Daily bars only for the rule: a crash inside a day is taken in full until the next close.
- Costs are modelled, not measured on real fills; cash earns nothing in the test.
- The backtest uses Binance prices; the forward record uses Kraken's. On the 520 overlapping days checked, the rule's on/off state was identical on both.
- Nothing here knows your situation. It is research, data and software output, not investment advice.
Risk. Crypto assets can lose most or all of their value. Backtested results are hypothetical and were produced with hindsight; past and simulated results do not predict future results. Only use money you can afford to lose.