Do trading indicators actually work?

We did not only test indicators. We ran a coin flip through the identical engine — same candles, same stops, same spread, same rules — and put its result in the column next to every indicator.

This is the part we would ask you to look at first, because without it none of the other numbers mean anything. If an indicator returns −0.1287R per trade, that is only informative once you know what entering at random returns on the same charts. An apparent edge that a coin flip also produces is not an edge; it is the cost of trading, or noise, wearing the name of an indicator.

It is also the strongest check available that the simulation is not cheating. A signal with no information in it should lose approximately the spread and nothing else. If random came out profitable, the engine would be reading the future and every other figure here would be worthless.

The detail: five-minute: −0.1287R against random’s −0.1074R. Fifteen-minute: −0.0524R against −0.0661R. Those gaps are 213 and 137 ten-thousandths of R across hundreds of thousands of trades. 55 indicators, collectively, add nothing you could not get by flipping a coin.

And the gradient is not about indicators at all. Random entries show the same curve — from −0.107R on five minutes to −0.0070R on daily. So the improvement as the bar slows is not indicators working better. It is that a slower bar carries less cost and less noise per trade, and a coin flip benefits from that exactly as much as an indicator does.

An earlier version of this page presented the gradient on its own and read it as evidence about indicators. The control shows that reading was wrong, and it is left here rather than quietly corrected because the mistake is the useful part: a number with no baseline beside it will be over-interpreted, including by the person who produced it.

On four-hour and daily charts the indicators start beating the control — by +0.0086R and +0.0104R respectively. Counted one indicator at a time rather than pooled, the picture is sharper still:

On five-minute charts 6 of 55 beat the control — about what chance alone produces from 55 coin flips. On four-hour charts 37 of 55 do, which chance does not produce easily.

But read the two columns together before getting excited. Beating the control is not the same as making money: on four-hour charts 37 of 55 still finish below zero. They lose less than a coin flip does, which is a real result and is not a profit. Daily is the strongest row in the table — 32 of 55 above zero — and it is also among the thinnest, with the fewest trades in the study.

Treat this as a hypothesis worth testing, not a conclusion. It is the most interesting result on the page and the least certain one, and those two things are related. The strategy tester runs the same engine if you want to push on it yourself.

• On fast charts there is no evidence of an indicator edge. Five- and fifteen-minute results are indistinguishable from random entries run through the same engine.
• The slower charts are the interesting part — and the least certain. On four-hour charts 37 of 55 indicators beat the control, on the smallest number of trades in the study. Interesting is not the same as proven.
• Win rate on its own tells you nothing about whether you make money. Several indicators here win most of their trades and still finish negative, because the losses are bigger than the wins.
• Costs decide fast trading. The spread is charged on every trade here, and on five-minute charts it is most of the result. Any backtest that leaves it out will look profitable and be wrong.
• Out-of-sample is the column to believe. It scores the last 30% of each window — data no setting was chosen against. Where it disagrees with the headline number, the headline is the one that is lying.
• No single indicator is a trading system. Nothing on this page was combined, filtered by session, or managed after entry. The first of those now has its own study: does making two indicators agree help? — every pair tested, each scored against the indicator that fires it.

Expectancy is the column that matters. A 70% win rate with a negative expectancy still loses money, and several indicators here manage exactly that. Profit factor and win rate are both easy to make look good by changing what counts as a trade; expectancy per trade is not.

Trade count is the credibility of the row. A profit factor of 3 from forty trades is noise wearing a suit. That is why the default filter is 200 trades and why thin rows are greyed rather than deleted — hiding them would be its own kind of dishonesty.

Entries fill at the next bar’s open, never the bar that produced the signal, and the stop is four median bars wide for that timeframe with the target at twice the stop. Scaling the stop to the bar is what makes the timeframes comparable: a fixed 20-pip stop is eight five-minute bars and less than one four-hour bar, so holding it constant measures the stop rather than the indicator.

It is not a ranking of which indicator to trade. Every one runs on its textbook settings — RSI 14, MACD 12/26/9, Bollinger 20/2 — and nothing else. Tuning each to its best parameters would produce a prettier table describing nothing but the tuning.

Then how does the signal service work?

A fair question, and the obvious one: if the laboratory says classic indicators do not clear their costs, why does the published record show what it shows?

Because this page tests something different. Every row above is a single indicator, alone, on its published settings, entering on every signal it gives, with a stop of four median bars and a target of twice that — in all market conditions, on all 15 instruments, forever. That is not how anybody trades, and it is not how our signals are produced. Ours combine several factors that must agree, decline far more setups than they take, and manage the trade after entry. Nothing on this page is a backtest of our signal engine, and it should not be read as one.

The one number that compares directly is expectancy, because it is in R either way:

The same standard, applied to us. Our record is 1,031 trades over about five months. The laboratory is 1,324,836 trades over up to five years — roughly 1285 times the sample. Every argument this page makes about small samples, out-of-sample testing and over-interpretation applies to our own results too, and the honest position is that five months is one market regime, not proof of an edge.

What we can say is narrower and checkable: every one of those trades was published before its outcome was known, the losses are still on the record, and the spread is deducted from the figure above exactly as it is from every row in this table. Judge it on that, and on how it looks in another year.

Citing this study

Check it yourself

The strategy tester uses the same engine, the same candles and the same spread model. Pick any row that interests you, change one number, and watch what the out-of-sample column does — that is the fastest lesson in curve fitting available for free. The live record is where our own trades are published before their outcome is known, with the losses left in.