Does trend following work on Bitcoin?

Yes, and you almost certainly cannot trade it.

That is not a dodge. Every trend and momentum rule we tested on daily and weekly crypto bars made money — some of them made spectacular amounts of money — and every single one was rejected, all for the same reason, and the reason was never trading costs. This page is about what that reason is and why it does not show up in the backtests people post online.

The chronological versions are in the research log, sweep #9.

The test

Three published rules, frozen at their original parameters, pointed at BTC, ETH and BNB. Two independent four-year windows: August 2018 to August 2022, and August 2022 to August 2026. Spot fees at 0.1% per side, stressed at 0.2%. The gates are the library's standard ones — net positive in both windows, still positive after deleting the single best trade, maximum drawdown around 20% or less.

The rules were Faber's 200-day moving-average timing model, Liu and Tsyvinski's one-week momentum rule from their 2021 crypto paper, and Carter's TTM squeeze breakout. Nothing was tuned. Parameters chosen by other people on other data are the cheapest honest hypotheses available, which is the whole point of testing them.

The results nobody disputes

Here is the 200-day timing rule — buy when the daily close crosses above the 200-day average, sell when it crosses back below — on Bitcoin's first window:

A profit factor above five. More than triple your money. This is the kind of table that sells a course, and on its own it is completely worthless.

Why it is worthless: one trade

Of that +315%, a single position — opened 29 April 2020, closed 19 May 2021 — contributed +317.8%. Remove that one trade and the window is negative. The other thirteen trades are whipsaw losses between −2.6% and −10.1%: buy the cross, watch it fail, sell the cross back, repeat, for a year, while waiting for the one move that pays for all of it.

Ethereum tells the identical story at a larger scale: +887.9% over the window, of which +889.9% came from one hold. Strip the best trade and you are at −2.0%.

This is precisely what the strip-best-trade check exists to catch, and it is the single most useful thing to run on any backtest you are shown. If deleting one trade changes the conclusion, you do not have a strategy. You have a story about one trade.

The second window is the trap

Run that same 200-day rule on 2022 to 2026 alone and you get:

Both windows pass every gate cleanly. A ten-point drawdown, a profit factor near seven, a hundred percent return. Anyone who backtested only the recent past would publish this in good conscience and would be wrong — not because the numbers are fake, but because the same rule falls apart in the window immediately before it.

Two independent windows is not a bureaucratic formality. It is the cheapest protection against the specific mistake of mistaking one regime for a law.

The result that is genuinely interesting

One-week momentum — long after an up week, flat after a down week, a rule with no parameters at all — is the strongest thing we found. Five of six asset-window cells are net positive, profit factors run 1.55 to 3.95, and they survive the strip-best test across about fifty trades per window. The effect Liu and Tsyvinski documented on 2011 to 2018 data is still visible in two windows that are entirely out of sample for it.

It is not in the library, and the reason is the only column that matters here:

Thirty to sixty-six points of drawdown against a bar of twenty. Monte-Carlo resampling of the same trades puts the 95th-percentile drawdown between 70 and 112 points. Raising fees from 0.1% to 0.2% per side changes none of it.

The actual answer

Crypto trend following works and it hands you the drawdown of crypto.

Buy-and-hold on these assets over these windows drew down 52 to 80 points. The trend systems cut that roughly in half — which is a real, measurable achievement, and half of seventy is still thirty-five. There is no version of holding a single volatile asset directionally, long enough to capture a trend, that also produces a twenty-percent drawdown.

That is why this library's crypto entries look the way they do. The three that cleared the bar either read positioning and exit within hours, or hold no direction at all:

• BTC funding-squeeze at 4h — in the market about 2% of the time.
• BTC funding-squeeze at 1h — about 2.3%.
• Funding harvest — market-neutral; it owns no direction to lose on.

The two walls

Put this next to the low-timeframe finding and the picture is complete in both directions.

Below the daily bar, crypto strategies die of costs. The edges are real and roughly five times smaller than the fees — the opening-range breakout is the clearest example.

Above the daily bar, they die of drawdown. Costs become irrelevant at four trades a year, and the thing that kills you is the asset itself.

Between those two walls sits a narrow band where something survives, and what survives there does not read price at all. That is not a slogan we started with. It is where eleven sweeps ended up.

What would change this answer

Not another sweep on candles. A trend system that clears a twenty-percent drawdown bar on crypto needs something that addresses drawdown structurally and is decided before the backtest: position sizing scaled to volatility, a basket rather than one asset, or an explicit decision to run it as a high-drawdown sleeve sized accordingly.

That last option is legitimate and this library will not pretend otherwise. Sweep #1 found a Donchian breakout on 4h bars that is positive and strip-best-surviving in all six cells at 26 to 76 points of drawdown. If you want crypto trend exposure, that is roughly the deal on offer: a genuine edge, at a drawdown most people discover they cannot sit through. Choosing it deliberately is a real decision. Discovering it in month four is not.