How Many Trades Before You Know a Signal Works? (2026)
Twenty trades tells you nothing. Measure a 54.6% strategy over twenty trades and the honest 95% range around your result runs from 32.8% to 76.4% — it comfortably contains both “broken” and “extraordinary”. To separate a 54.6% edge from a coin flip with any confidence you need roughly 962 trades. Almost nobody waits that long before deciding, which is why almost every verdict on a signal service — positive or negative — is a verdict on noise.
Disclosure: FuturesAI is our own product and the measured figures here are ours. This article argues that our own top confidence tier cannot be claimed from, which is the conclusion the arithmetic forces.
Jump to a Section
- Why Twenty Trades Is Not Evidence
- How Many You Actually Need
- The Cruel Part: Smaller Edges Need More
- Nine Losses in a Row Is Normal
- What This Says About Our Own Data
- What to Do Instead of Waiting
Why Twenty Trades Is Not Evidence
Take a strategy whose true win rate is 54.6% — a real, modest edge. Run it a given number of times and ask how wide the range of plausible results is.
| Trades observed | 95% range around the measurement | Does it exclude a coin flip? |
|---|---|---|
| 20 | 32.8% – 76.4% | No |
| 100 | 44.8% – 64.4% | No |
| 776 | 51.1% – 58.1% | Yes |
Read the first row again. After twenty trades, a genuinely good strategy can legitimately show you anything from a disaster to a miracle, and you have no way to tell which one you are looking at. The result is dominated by chance, not by quality.
This cuts both ways, and the second direction is the one that costs people money. A service with no edge at all can easily produce 13 wins out of 20 — that is well inside the range of a pure coin flip. Someone watching that run will conclude it works, size up, and then meet the average.
How Many You Actually Need
The number of trades required to say “this is not a coin flip” with conventional confidence depends entirely on how large the edge is.
| True win rate | Trades needed to distinguish it from 50% |
|---|---|
| 52% | 4,898 |
| 54.5% | 962 |
| 55% | 778 |
| 60% | 189 |
| 65% | 80 |
| 70% | 42 |
At a five-day holding period, 962 trades is years of trading. That is the honest timescale on which a modest edge becomes provable, and it is wildly out of step with how people actually evaluate — a few weeks, a few dozen trades, then a verdict.
The Cruel Part: Smaller Edges Need More
Look at the shape of that table. Going from a 60% edge to a 55% edge does not double the trades required — it quadruples them. Down at 52% you need nearly five thousand.
The requirement scales with the square of how small the edge is, so the edges most likely to be real in markets are precisely the ones hardest to confirm. Big edges are easy to verify and rare. Small edges are common and take years to demonstrate.
This has an uncomfortable implication for marketing. A service advertising 90% accuracy is making a claim that would be easy to verify on a small sample — 42 trades would do it at 70%, fewer at 90%. So ask for the sample. If a service claims a very high win rate and cannot produce a few hundred trades to support it, the claim and the evidence are not the same size.
Nine Losses in a Row Is Normal
The other thing a small sample hides is how ugly a winning strategy looks along the way.
At a 54.6% win rate, losses come up 45.4% of the time, and streaks are unavoidable. The longest losing run you should expect to see:
| Over this many trades | Expect a losing streak of about |
|---|---|
| 50 | 5 in a row |
| 100 | 6 in a row |
| 500 | 8 in a row |
| 1,000 | 9 in a row |
Nine consecutive losses is not evidence that something broke. On a real edge, over a thousand trades, it is the expected outcome. And it will arrive at some point, which means the decision to abandon a strategy during a losing run is a decision made at exactly the moment the data is least informative.
Worth pairing with the arithmetic from the break-even side: a modest win rate only pays if your targets are wide enough to outrun costs. A strategy that is both thin-edged and short-target is one where streaks hurt most and confirmation takes longest.
What This Says About Our Own Data
We publish the scoring data, so this applies to us first.
Across 26,630 scored samples, the working band won 54.6% of the time. That band has tens of thousands of samples behind it, which is why we are willing to describe it at all. But the distribution is not even. The top confidence tiers hold almost nothing: 215 samples in the second-highest, 8 in the next, and zero in the highest.
Eight samples is one winner away from any story you like. Run the interval calculation on eight observations and it spans nearly the whole range between failure and brilliance. So the honest position is that our own highest-confidence tier supports no claim whatsoever — not a good one, not a bad one. It is empty of information, and we say so rather than quoting its percentage.
That is also the specific thing to check when anyone advertises the accuracy of a “high confidence” tier. The tier is where sample sizes are smallest by construction, because high scores are rarer. It is the part of the data most likely to be noise and the part most likely to be quoted.
What to Do Instead of Waiting
Nobody is going to run a thousand trades before deciding anything. So the practical question is how to decide well on less.
- Ask for the sample size next to every percentage. A win rate without an n is not a measurement. This single question disqualifies most claims.
- Judge the process, not the last ten trades. Was every signal taken? Was the cost assumption realistic? Was the horizon the one being marketed? Those are answerable now.
- Pool across instruments. Twenty trades on one ticker is noise; twenty trades across ten tickers accumulates evidence faster, provided the strategy is supposed to be general.
- Size for the streak, not the average. If nine consecutive losses would end you, the edge is irrelevant — you will not be there when it pays.
- Use the measurement you already have. Published historical data on thousands of samples is worth more than your own first month, and it costs nothing to check.
The scoring data used here, including the empty top tier, is public at FuturesAI if you want to run these intervals yourself.
🔗 Related
FAQ
Where do these numbers come from?
Standard sample-size and confidence-interval arithmetic for a proportion, at a 95% two-sided level with 80% power. Any statistics package reproduces them; nothing here is proprietary.
Does a longer holding period reduce the trades needed?
No. The requirement is in trades, not in time. A longer horizon means the same number of trades takes longer to accumulate, which makes confirmation slower in calendar terms even though the edge may be larger.
What if a service shows me a long backtest instead?
Better than twenty live trades, with different failure modes — a backtest can be fitted to its own history. Ask whether the rules were fixed before the test period and whether costs and slippage were included.
My strategy is 8 for 10. Should I scale up?
Ten trades cannot distinguish an 80% strategy from a lucky coin flip. Scaling on that basis is sizing on noise, which is how a normal losing streak becomes an account-ending one.
Sources
- FuturesAI public scoring endpoint — 26,630 samples across 59 tickers, 54.6% in the working band; 215 / 8 / 0 samples in the top three tiers. Retrieved September 25, 2026.
- Sample-size and interval figures computed with standard proportion arithmetic (95% two-sided, 80% power) and reproducible from the win rates shown.
- futuresai.io — public scoring dashboard.
This article is for informational purposes only and is not investment advice. Trading carries risk, and leveraged trading in particular can result in the loss of your entire principal. Historical measurements do not predict future results.
