FuturesAI Scoreboard: Do UP Calls Beat DOWN Calls?
Here is a filter you can test in one sitting: ignore every UP call, or ignore every DOWN call. FuturesAI publishes both groups. So you can check whether either one earns the right to be followed alone.
See the FuturesAI payback rates
What This Tests
- What does the scoreboard say about UP and DOWN calls?
- How wide is the noise on each side?
- Is a 1.9-point gap a real difference?
- How many calls would it take to tell?
- A direction-filter rule you can keep
What does the scoreboard say about UP and DOWN calls?
FuturesAI’s public scoreboard, with its last automatic check dated November 11, 2026, lists UP calls at 45.3% accuracy (96 of 212) and DOWN calls at 47.2% (59 of 125) (futuresai.io/en/quant/scoreboard, read November 14, 2026). The page notes that both directions are published so the record can’t be framed one-sidedly.
When I checked the page, I noticed that the direction block adds up to 337 scored calls (155 correct), while the headline accuracy box counts 322 predictions (148 correct, 46.0%). The two blocks are cut at different points. We unpacked a similar mismatch in FuturesAI Scoreboard: How to Reconcile 317 vs 332 Calls.
For this test I use the direction block as printed. Its own total is the fair denominator.

How wide is the noise on each side?
A hit rate measured on a few hundred calls carries a rough 95% error band of 1.96 times the square root of p(1 − p) divided by n. For the UP calls that band runs from 38.6% to 52.0%, and for the DOWN calls from 38.4% to 56.0%.
Work UP through by hand. p = 96 ÷ 212 = 0.4528. The standard error is the square root of (0.4528 × 0.5472 ÷ 212), which is 0.0342. Times 1.96 is 6.7 points either side of 45.3%.
| Group | Hits / calls | Hit rate | Rough 95% band |
|---|---|---|---|
| UP calls | 96 / 212 | 45.3% | 38.6% to 52.0% |
| DOWN calls | 59 / 125 | 47.2% | 38.4% to 56.0% |
| Both together | 155 / 337 | 46.0% | 40.7% to 51.3% |
Counts from the FuturesAI scoreboard, read November 14, 2026. Bands calculated by this site with the normal approximation; a rough guide, not an exact interval.
Look at the overlap. The two bands share almost their entire range.
Is a 1.9-point gap a real difference?
A 1.9-point gap between DOWN (47.2%) and UP (45.3%) hit rates is not a real difference at these sample sizes, because the noise on the gap itself is about 11 points. The standard error of the difference is the square root of (0.0342² + 0.0447²), or 5.6 points, and 1.9 ÷ 5.6 is a z-score of 0.34. A gap would need a z-score near 2 to look meaningful.
In plain words: if both sides were exactly equally accurate, a gap this large would show up by luck far more often than not.
I found that this is the point where most people go wrong with a record like this. They see 47.2% beside 45.3% and read it as a ranking.
How many calls would it take to tell?
To confirm a gap of 1.9 points between two hit rates near 46%, you would need roughly 5,000 calls on each side. The estimate comes from n = 2 × p(1 − p) × (1.96 ÷ gap)² with p = 0.46 and gap = 0.019.
FuturesAI’s record has 337 scored calls in total. That is about 15 times too few for this question.
Does that mean the scoreboard is useless? No. It means it can answer coarse questions (is the overall record near a coin flip?) but not fine ones (does one direction beat the other?). The overall band in the table, 40.7% to 51.3%, is wide for the same reason. See FuturesAI Scoreboard: How Wide Is the Error Band? for the same logic on the headline number.
A direction-filter rule you can keep
Before you filter any signal source by direction, market or tag, write down the group size and compute the rough band. If the bands for the two groups overlap by more than half, treat the split as noise and keep it out of your rules.
- Copy the counts: hits and total for each group, from the page itself.
- Compute the rate: hits ÷ total.
- Compute the band: 1.96 × the square root of p(1 − p) ÷ n.
- Check the overlap: do the two bands share most of their range?
- Decide: overlapping bands mean no filter, and a re-check once the sample has grown a lot.
The same habit applies to the number tag tested in FuturesAI Ledger: Does a Higher Call Number Mean More Hits? Different split, same arithmetic.
Tips
- Screenshot the counts with the date. The page updates, and old numbers vanish.
- Test one filter at a time. Each extra split cuts the sample you have left.
- Use the scoreboard to check claims, not to pick trades.
Warnings
- This is arithmetic on a short public record, not a forecast and not investment advice.
- FuturesAI’s own page says its AI signals are observation scores, not buy or sell recommendations.
- Leveraged futures can lose the full principal. Past hit rates do not guarantee future results.
FAQ
Are UP calls or DOWN calls more accurate on FuturesAI?
DOWN calls show 47.2% (59 of 125) and UP calls 45.3% (96 of 212) as of November 14, 2026. The gap is far smaller than the noise on a sample this size, so neither side can be called better.
Does the scoreboard include trading fees?
No. FuturesAI’s scoreboard states that its returns are bought at the price when the prediction was made and sold at the scoring-day price, with no fees. Fees would lower any traded result.
Why do the scoreboard totals not match each other?
Different blocks of the page count the ledger at different cut-offs. Here the direction block totals 337 calls while the headline accuracy box counts 322. Use each block’s own total when you compute rates from it.
See the FuturesAI payback rates
