In development / Explainer
Strategy Creator · Explainer

Why 9 out of 10 trading strategies fail

Of the 227 strategy variants we have tested so far, 89% failed. It isn’t bad luck, nor a problem on our side: this is what usually happens when you test honestly. Below we explain, step by step, why — and what to watch out for when someone shows you a strategy or a trading robot that “surely works”.

Reading time: ~8 minutes · October 8, 2026

1. Luck looks exactly like skill

Picture 200 people flipping a coin every day: heads means +1%, tails −1%. Nobody has any edge. After a year, some will be up and some down — and the best of them will look like a genius. We simulated it:

200 coin-flipping “strategies” (+1% or −1% a day, at random).

That is exactly what happens when you try many strategies on historical data. If you test 200 ideas, a few will look excellent purely by chance. The problem isn’t that good strategies don’t exist — the problem is that, looking only at the past, you can’t tell whether one is good or just lucky.

What to watch out for

Ask how many variants were tried before arriving at the one presented. A good result found in 3 attempts means something; one found in 300 means almost nothing.

2. Tweaking until it looks good

The second trap is subtler. You have an idea, you test it, it comes out weak. You change a rule a little — it comes out better. You change it again, and again. After 20 tweaks, the strategy looks flawless on the past. But every tweak “learned” the noise in that data, not a real market mechanism. The strategy describes the past perfectly and nothing of the future.

This is called overfitting. The typical sign: very precise rules, like “buy when the indicator crosses 37.5, but only on Tuesdays and Thursdays”, with no explanation of why it should work.

How we protect ourselves

Rules, costs and success criteria are written and sealed with a digital fingerprint before the test, so they can no longer be adjusted after the result. A period of data stays untouched and is used only once, at the end.

3. Costs eat the difference

Many market patterns really exist — they are just smaller than the cost of using them. On every trade you pay the difference between the buying and the selling price (the spread), the commission and, often, a fill at a slightly worse price than expected.

Illustrative example, not a real strategy’s numbers: an edge of 0.10% per trade looks good, but after costs almost nothing is left, and at double costs it becomes a loss.

A gain of a few hundredths of a percent per trade disappears after costs. And the costs assumed in a test are almost always lower than the real ones. A related trap: the reverse of a losing strategy doesn’t become a winner — costs are paid in both directions.

How we protect ourselves

We measure costs on real fills instead of assuming them, and we require a strategy to stay positive even at double costs.

4. Information from the future

The most common technical error: the test uses, without your noticing, information that didn’t yet exist at the moment of the decision. A few examples:

Errors like these produce spectacular results. That’s why, for us, a result that looks too good on a simple idea is treated as an error until proven otherwise.

How we protect ourselves

Every simulated decision uses only what was known at that moment — including the list of companies and the data as they were published at the time.

5. The market changes

Even a real effect doesn’t last forever. When a pattern becomes known, more participants exploit it and the edge shrinks or disappears. The rules, the technology and the people in the market change too. A strategy that worked brilliantly for a few years may stop working entirely in the following years — we had very good ideas on the build years that vanished completely on the unseen period.

How we protect ourselves

The period kept untouched is the most recent one, and a strategy must be positive in most years, not only in a lucky stretch.

6. Too few trades

20 winning trades out of 30 can look convincing, but statistically they prove almost nothing: an equally good result often appears by pure chance. You need hundreds of trades, across different market conditions, to separate edge from luck.

How we protect ourselves

We require a statistically significant result and compare every strategy with a control test: the same rules, but with entry moments chosen at random. If it doesn’t clearly beat it, it has no edge.

7. Simulation isn’t reality

Even a strategy that passes every test on historical data can behave differently in reality: orders are filled later or only partly, at certain hours there are fewer buyers and sellers, the connection has delays. That’s why, before real money, strategies run for months on virtual accounts, day by day, and must reproduce the simulation exactly.

And this forward test can’t prove in a few months that a strategy is good — it can only reject it in time. Good results aren’t taken as proof; bad ones stop the strategy right away.

In short: the traps and the defences

The trapHow we protect ourselves
Luck mistaken for skillWe count every variant tried; we require a statistically significant result
Tweaking until it looks goodRules sealed before the test; unseen data used only once
Costs ignoredCosts measured on real fills; the strategy must survive double costs
Information from the futureEvery decision uses only what was known at that moment
The market changesThe unseen period is the most recent one; positive in most years
Too few tradesA control test with random entries
Simulation isn’t realityA forward test on virtual accounts; real money only at the end, and a little

What to watch out for when someone shows you a strategy or a robot

Whether it’s a course, a group, a “robot that makes money on its own” or an offer to manage your money, these signs should make you stop and ask more:

What this means for you

The fact that 9 out of 10 ideas fail isn’t a bad sign but a sign that the tests are honest. Research in which every idea comes out a “winner” either doesn’t test enough or is fooling itself. We’d rather reject an idea in the lab than test it with someone’s money.

Educational material. This is not investment advice or an offer. The numerical examples are illustrative, and the simulation uses randomly generated data. Past results, simulated or real, do not guarantee future results.