Trend following wins 36% of the time. That is the point.
The rule we ran on a live server lost on almost two trades out of three in nine years of replay. Its profit came from a handful of long trends. Here is what that looks like from the inside.
How this was made: the code, the research runs and the first drafts of this text were produced by an AI assistant (Claude) working under the operator's direction; the operator made every decision and checked the numbers. All trading was on paper: real prices, no real money. Nothing here is investment advice.
The rule
On 4-hour bars of the eight most traded coins on Upbit: buy when the close breaks the highest close of the last 55 bars; sell when it falls two average true ranges below its highest close since entry; buy nothing while Bitcoin is below its 100-bar average; cut any position that loses 5%. The replay ran from January 2019 with costs included.
The trade statistics
| Nine years, 1,037 trades | |
|---|---|
| Win rate | 36.4% |
| Average losing trade | −3.37% (median −2.78%, held about 1.2 days) |
| Average winning trade | +12.98% (median +5.03%, held about 4.5 days) |
| Profit from the best 10% of trades | 211% of the total net profit |
| Longest losing streak | 16 trades |
| Four losses in a row | 246 times |
The best tenth of the trades made more than twice the rule's total profit. The other nine tenths, taken together, lost money. Miss a few of the big trends and the whole result changes, which is why a trend system that is offline at the wrong moment can turn a winning year into a losing one.
How often a stretch of time loses
| Holding period | Chance it ended in a loss |
|---|---|
| 30 days | 46% |
| 90 days | 29% |
| 180 days | 17% |
| 365 days | 5% |
Almost half of all 30-day stretches ended in a loss, even for a rule that made money over the nine years. Losing streaks also told us nothing about what came next: after four straight losses, the following 90 days were positive 76% of the time, against 71% at any time.
Why not exit faster?
It is tempting to watch prices in real time and sell the moment the trailing line is crossed. In the replay, 456 of the 1,037 trades crossed the line during a bar but closed back above it, and 228 of those went on to finish with a profit. Acting on every intrabar cross would have cut the final balance by 75%. Noise inside the bar shakes out the trades that pay for everything else.
What this means for anyone running one
- Expect to lose most trades and most weeks. Our two-week live test lost on five trades out of five, which this rule does regularly. The server test.
- Judge it over years, not weeks: about 34 months of trades to separate skill from luck.
- Keep it running. Its returns are concentrated in a few trends you cannot predict.
The replay made +92% a year when settings were chosen from past data only; after allowing for the biases we knew about, we expected +28% to +46% a year with drawdowns of 27% to 33%. Even that turned out to be a backtest number, not a live one.