Do AI Trading Signals Actually Work? We Measured 26,630 of Ours (2026)
We run an AI scoring system for equities at FuturesAI, and we publish the scores openly. So we pulled our own numbers and checked what they are actually worth. Across 26,630 scored samples on 59 tickers, the low-score bucket won 42.6% of the time and the mid-to-upper buckets won 54.6% — a 12 percentage point spread. That spread is the whole story, and it points the opposite way from how signal services usually sell themselves. The reliable finding is not “the high scores are winners.” It is “the low scores are losers.” Two other numbers matter more than the headline: the top confidence bucket has 8 samples, which is not enough to conclude anything, and on the one ticker we backtested, simply holding the stock beat the signal strategy by a wide margin.
Disclosure: FuturesAI is our own product. The figures below come from its public API and you can pull them yourself — the endpoints are named in the Sources section. We have published the unflattering numbers alongside the good ones for exactly that reason.
Jump to a Section
- What Did We Actually Measure, and on What?
- What Do the Score Buckets Show?
- Why Is the “Highest Confidence” Bucket the One to Ignore?
- Does Holding Longer Change the Answer?
- What Number Does Signal Marketing Usually Leave Out?
- So What Is an AI Signal Actually Good For?
- How Do You Check Any AI Signal Service Before Paying?
What Did We Actually Measure, and on What?
The system scores equities and sorts every scored sample into ten buckets, from the weakest tenth to the strongest tenth. For each bucket we track how often the position finished higher after a fixed holding period, and by how much on average.
The snapshot in this article was computed on September 25, 2026. It covers 26,630 scored samples across 59 tickers, measured over a 20-trading-day holding period. That last detail matters more than it looks, and we come back to it in section four.
Three limits are worth stating before the numbers, because they bound what you can conclude:
- This is equities, not crypto perpetuals. Different asset, different volatility, different holding economics.
- Fifty-nine tickers is a real sample but not a market. Sector concentration could be doing work we cannot see from the aggregate.
- The measurement window sits inside a period when the underlying market rose. Some of the win rate is the market, not the model. We could not separate those two with the data we have, and we say so again in section five rather than quietly dropping it.
What Do the Score Buckets Show?
Here is the full table as the API returned it. Empty buckets are included rather than hidden, because the gaps are part of the finding.
| Score bucket | Samples | Win rate | Avg return |
|---|---|---|---|
| D1 (weakest) | 0 | — | — |
| D2 | 11 | 45.5% | +6.26% |
| D3 | 687 | 42.6% | −0.04% |
| D4 | 5,120 | 55.9% | +3.42% |
| D5 | 10,132 | 55.2% | +2.92% |
| D6 | 8,019 | 52.8% | +3.45% |
| D7 | 2,438 | 55.5% | +4.05% |
| D8 | 215 | 57.2% | +3.28% |
| D9 | 8 | 12.5% | −5.23% |
| D10 (strongest) | 0 | — | — |
Group the buckets that actually hold data and the picture is clean. The low band (D2–D3, 698 samples) finished ahead 42.6% of the time. The working band (D4–D8, 25,924 samples) finished ahead 54.6% of the time with an average return of +3.29%. The gap is 12.0 percentage points.
Now read that a second time, because the useful half is the left-hand side. A 54.6% win rate is a modest edge that fees and timing can erode. A 42.6% win rate is a reliable way to lose money slowly. The model is far better at telling you what to skip than at telling you what to buy.
Why Is the “Highest Confidence” Bucket the One to Ignore?
Look at D8 through D10 again. D8 has 215 samples. D9 has eight. D10 has none at all.
D9’s win rate reads 12.5%, which looks alarming until you realise it means one winner out of eight. At that sample size the number carries no information in either direction. If it had come out at 87.5% we would be equally unable to use it, and a less careful write-up would have led with it.
This is the single most common trick in AI signal marketing. A service advertises the win rate of its top-confidence tier without saying how many trades that tier contains. Ninety percent accuracy on twelve trades is a coin that landed heads a few times. Our own top tier is empty enough that we cannot make a confidence claim from it, so we do not make one.
The scores also do not climb cleanly. D4 is 55.9%, D6 drops to 52.8%, D7 climbs back to 55.5%. If the score were a true confidence dial, each step up would beat the one below it. It does not. That kills any copy of the form “a score of 90 means high conviction” — the data simply does not support reading the number that finely. What it supports is a coarse split: bottom band bad, middle band mildly positive.
Does Holding Longer Change the Answer?
Yes, and this is where most people using signals go wrong. We ran the backtest endpoint against a single large-cap ticker with a 20 basis point round-trip fee applied, and varied only the holding period.
| Holding period | Trades | Win rate | Avg return per trade |
|---|---|---|---|
| 1 day | 58 | 53.4% | +0.41% |
| 5 days | 30 | 70.0% | +2.83% |
At a one-day horizon the signal is barely distinguishable from a coin flip, and a +0.41% average per trade is thin enough that fees and slippage can eat most of it. Stretch to five days and both the hit rate and the size of the average win improve materially.
The practical reading: the shorter the horizon, the weaker the signal. Which is awkward, because “should I buy today” is exactly the question most people bring to a signal service. The honest answer from our own data is that the tool is worst at the timeframe it is most often asked about.
One caveat we will not bury: this holding-period table is a single ticker. It is directionally consistent with the bucket data, but it is not a market-wide result and should not be read as one.
What Number Does Signal Marketing Usually Leave Out?
The benchmark. On that same ticker over that same window, the strategy returned +116% in total at the five-day horizon. Simply buying the stock and holding it returned +410%.
The signal had the better win rate. Buy-and-hold had the better outcome. Those two things are not in conflict — a strategy that is out of the market between trades gives up the compounding that a holder keeps, and a rising market rewards the holder for doing nothing.
This is why a win rate quoted with no benchmark is close to meaningless. A 70% hit rate sounds like an edge until you learn that the boring alternative made three times as much over the same period. Any service showing you accuracy without showing you what you would have earned by not using it is showing you the flattering half of the picture.
So What Is an AI Signal Actually Good For?
On this evidence, one thing: filtering.
The finding that survives every caveat in this article is the low band. 42.6% over 698 samples is a persistent, unhelpful place to be, and avoiding it is an action you can take without predicting anything. You are not being told what will go up. You are being told which candidates have historically been a worse-than-even bet, so you can stop considering them.
That is a smaller claim than “AI picks winners,” and it is the only one the numbers carry. It also happens to be the more useful claim for anyone whose main problem is losing money on positions they should never have opened — which, for most people trading actively, is the actual problem.
If you want to check any of this, the scoring data and the backtest endpoint are public at FuturesAI. Pull the same endpoints we did and you should reproduce these tables.
How Do You Check Any AI Signal Service Before Paying?
Five questions. If a service cannot answer them from its own published data, that is the answer.
- How many trades are in the tier you are advertising? Accuracy on a tier with fewer than a few hundred samples is noise. Ask for the count next to every percentage.
- What holding period produced that number? A win rate is meaningless without it, and services rarely volunteer that their strong figure comes from a horizon far longer than the one they market to day traders.
- What did buy-and-hold return over the same window? If the benchmark is missing, assume it is unflattering.
- Are fees included? A strategy that trades often can look profitable gross and lose money net. Our tables above apply a 20 basis point round trip; many published backtests apply nothing.
- Does accuracy rise monotonically with confidence? If the “90 score” tier does not reliably beat the “70 score” tier, the confidence number is decorative.
Those five questions will disqualify most of what is sold as AI trading intelligence, including, on the strict reading, the conclusion-style version of our own product. We would rather say that here than have you discover it after paying someone.
🔗 What this costs you
A thin edge only survives if your costs are thinner. If you are trading often enough for signals to matter, the fee side is where the arithmetic is actually decided.
FAQ
Does a 54.6% win rate mean the signal is profitable?
Not on its own. Win rate says nothing about the size of wins versus losses, and nothing about what you would have made doing something simpler. In our own five-day test the strategy won more often than it lost and still finished well behind buy-and-hold.
Why publish numbers that make your own product look worse?
Because the alternative is a claim we cannot support. The data supports “avoid the bottom band.” It does not support “follow the top score,” so we are not selling that, and the top band does not have the sample size to sell it with.
Is the top-score tier really only eight trades?
Eight in D9 and zero in D10 as of the September 25, 2026 snapshot. Those buckets fill slowly by design, since a high score requires several conditions to line up at once. Until they hold several hundred samples each, no confidence claim should be read from them.
Does this apply to crypto futures too?
These specific figures are equities. Crypto perps have different volatility and funding costs, and a shorter effective holding period, all of which tend to make a thin edge thinner rather than thicker. Do not carry these numbers across without re-measuring.
What is the single most useful takeaway?
Treat the score as a reason to remove candidates rather than a reason to add them. That is the part of the data that held up under every caveat in this article.
Sources
- FuturesAI score decile endpoint,
/api/quant/score-deciles— snapshot computed 2026-09-25, 26,630 samples, 59 tickers, 20-day horizon. - FuturesAI backtest endpoint,
/api/quant/backtest— single-ticker runs at 1-day and 5-day horizons with a 20 bp round-trip fee, retrieved 2026-09-25. - 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. Past performance does not predict future results, and every figure above is a historical measurement over a specific window that will not repeat. Figures retrieved from the FuturesAI public API on September 25, 2026 and may change as new data is scored — confirm current values at the source before relying on them.
