DRAWUP

the method

What is backtesting?

Backtesting means running a trading rule over prices that have already happened and recording what it would have done. It is easy to do and easy to do in a way that produces a number with no predictive content at all. We have run 2,160 of them and published every one, including the failures.

the short answer

four things that separate a test from a story

Decide on one bar, trade on the next. If a rule uses today's closing price it cannot also trade at today's close, because at the moment of the decision that price did not exist yet. Getting this wrong is the commonest way a backtest invents money, and it is invisible in the result.

Pay the costs. Every change of position costs something. We charge 5 basis points a side on every trade. A rule that trades daily and one that trades twice a year look very different once they both pay.

Split the history and judge on the half you did not look at. We cut at 70%. Anything chosen, tuned or noticed on the first part has to prove itself on the second.

Publish the rules that failed. A single strategy with a good chart tells you nothing, because you cannot see how many were tried. We run the same 15 rules on every market and show all of them.

what we found

how the 15 rules actually did

All of it comes from 144 markets, 124 crypto and 20 currencies, indices and commodities, each run through the same 15 rules on daily bars with 5 basis points of fees a side. Every figure below is from the out of sample half, the part no rule was chosen on. Holding lost a median -47% over those windows, which was a hard period, and that colours anything here that compares a rule to holding.

The most striking thing in the table is not which rule won. It is that the win rates are mostly low, that the medians are mostly negative, and that the rule with the best record (rsi(2) mean reversion, ahead on 68% of markets) is a short holding period mean reversion rule rather than any of the trend rules people usually mean by "technical analysis".

every rule, out of sample, across 144 markets
rulemade money onmedian returnmedian worst fall
rsi(2) mean reversion98 of 144+8.9%-31%
bollinger(20,2) mean reversion58 of 144-3.8%-37%
200 day trend, 10% stop45 of 144-15%-31%
rsi(14) mean reversion40 of 144-20%-45%
55 day breakout40 of 144-14%-34%
20 day breakout38 of 144-22%-47%
10/50 moving average38 of 144-15%-43%
macd(12,26,9)37 of 144-24%-53%
buy and hold33 of 144-47%-68%
50/200 moving average33 of 144-23%-40%
3 month momentum32 of 144-21%-44%
6 month momentum30 of 144-28%-43%
above the 200 day average29 of 144-22%-40%
12 month momentum21 of 144-6.6%-25%
dual momentum20 of 144-7.1%-22%

the catch

the result you get depends on when you looked

Our out of sample windows end between late 2025 and November 2026, and holding lost a median -47% over them. In a period like that almost any rule that spends time out of the market will beat holding, and 70% of ours did. That is a fact about the period, not a discovery about the rules, and a page that quoted the 70% without the -47% would be misleading you with true numbers.

This is why every market page here shows both halves, every rule, and the dates. You can read the test rather than trust the summary.

questions

what people ask

What is backtesting?

Running a trading rule over historical prices to record what it would have done. Done properly it includes trading costs, never uses information that was not available at the moment of the decision, and is judged on data the rule was not chosen on.

Is backtesting reliable?

Only as far as its method. Across 144 markets, the rule that looked best on the first 70% of a market's history went on to make money in the remaining 30% only 32% of the time.

How much history do you need to backtest?

Enough that the slowest rule has something to work with and enough left over to test on. We require 700 daily bars, because a 200 day moving average needs 200 of them before it has a value at all.

What trading costs should a backtest include?

At minimum a cost on every change of position. We use 5 basis points a side. Funding, borrow and slippage are additional and we say on every page that we do not model them.

the evidence

every figure on this page comes from these

144 markets, 15 rules, 2,160 tests, every one published with both halves and the dates. Read them at the market pages. Figures on this page were computed 2026-11-17.

we publish the rules that failed too

more

the rest of the explainers