FuturesAI Signals: Why Slicing 100 Calls Finds Fake Patterns
Short answer. On FuturesAI’s public ledger, read on November 18, 2026, calls with a 5 to 6 day window hit 6 of 23 (26.1%). Calls with a 20 to 22 day window hit 41 of 77 (53.2%).
That gap looks like a filter. It is mostly crypto showing up in the short window, and crypto calls were almost all UP.
Not investment advice. FuturesAI makes no profit promise, and leveraged futures can lose the full principal.
Cut a list of 100 outcomes into enough slices and one of them will look impressive. That’s arithmetic, not insight. This post cuts the ledger four ways and then shows what the cuts are worth.
See the FuturesAI payback page
Jump to a Section
- What does the ledger show when you slice it four ways?
- Is window length a filter, or is it just crypto?
- How often does a slice look good by chance?
- Why do the UP and DOWN gaps disagree?
- What would a fair filter test look like?
What does the ledger show when you slice it four ways?
FuturesAI’s scoreboard ledger lists 100 scored records, 47 hits and 53 misses as read on November 18, 2026, and slicing those 100 by window, direction or market gives hit rates from 25.0% to 62.8%. Every slice is smaller than 100, so each one carries a wider margin of error than the whole.
When I checked the 100 rows against the page header, I found that my hit count came to exactly 47, so the tallies below use the same records the page shows. Each row lists a frozen price, a resolved price and the window dates.
| Slice of the 100-record ledger | Hits | Hit rate | Rough 95% range |
|---|---|---|---|
| All 100 records | 47 of 100 | 47.0% | 38% to 57% |
| Window 5 to 6 days | 6 of 23 | 26.1% | 13% to 46% |
| Window 20 to 22 days | 41 of 77 | 53.2% | 42% to 64% |
| UP calls | 20 of 57 | 35.1% | 24% to 48% |
| DOWN calls | 27 of 43 | 62.8% | 48% to 76% |
| Crypto | 6 of 24 | 25.0% | 12% to 45% |
| Korean stocks | 24 of 46 | 52.2% | 38% to 66% |
| US stocks | 17 of 30 | 56.7% | 39% to 73% |
The last column is a rough 95% range for the true hit rate behind each count. Read it as a sanity check on how far a bucket could be from the truth, not as a forecast.
Notice how wide those ranges are. The short-window range (13% to 46%) and the long-window range (42% to 64%) overlap by only a few points. UP (24% to 48%) and DOWN (48% to 76%) just touch. Pairs that far apart look like a real difference, which is exactly why the next sections check whether they are.
Is window length a filter, or is it just crypto?
On FuturesAI’s ledger, window length cannot be separated from market: 23 of the 24 crypto records have a 5 to 6 day window, and all 76 stock records have a window of 20 to 22 days. So the short-window hit rate of 26.1% is really the crypto hit rate described another way.
| Market | Window | UP calls (hits of calls) | DOWN calls (hits of calls) |
|---|---|---|---|
| Crypto | 5 to 6 days | 5 of 22 | 1 of 1 |
| Crypto | 20 to 22 days | none | 0 of 1 |
| Korean stocks | 20 to 22 days | 7 of 20 | 17 of 26 |
| US stocks | 20 to 22 days | 8 of 15 | 9 of 15 |
Check the join by hand. Crypto 5 + 1 = 6 hits in the short window. Korean stocks 7 + 17 = 24 hits and US stocks 8 + 9 = 17 hits, which with the one crypto DOWN miss gives 41 long-window hits of 77. Add 6 and you land on 47.
Now look at direction inside crypto. 22 of its 23 short-window calls were UP, and they hit 5 times. So the short-window label, the crypto label and the UP label all point at about the same 22 calls.
That’s what I mean by a confound. A filter on window length is a filter on three things at once, and you cannot tell which one is doing the work from this table.
How often does a slice look good by chance?
If every call on the ledger had a flat 47% chance of a hit, the probability of seeing 6 or fewer hits in a bucket of 23 would be about 3.4%, and the chance that at least one of 10 independent slices would show a result that extreme would be about 29%. Slices of one ledger are not fully independent, so treat 29% as a rough guide.
The arithmetic is short. The probability of a result that extreme in one slice is 0.034, so the chance of avoiding it in one slice is 0.966. Across 10 slices that is 0.966 to the 10th power, or 0.71, and 1 minus 0.71 is 0.29.
Honestly, I found that part uncomfortable. A 3.4% result sounds rare until you remember how many ways you sliced the data to find it. Each extra cut is another lottery ticket.
A bucket that stands out is a question to test, not an answer to follow.

Why do the UP and DOWN gaps disagree?
The same scoreboard page shows two different UP versus DOWN pictures: the 100-record ledger has UP calls at 35.1% (20 of 57) and DOWN calls at 62.8% (27 of 43), while the page’s all-time direction summary read the same day shows UP at 44.4% (96 of 216) and DOWN at 48.0% (61 of 127). The gap is 27.7 points in one and 3.6 points in the other.
Both come from one public page, so neither is wrong. The ledger holds 100 records, and 22 of them are crypto UP calls that hit only 5 times. The longer summary spreads that stretch across a much bigger pile.
When I compared the two figures, I noticed that the bigger sample shrinks the gap by more than 20 points. That’s the usual pattern when a gap comes from a small sample. For the fuller version of the direction question, see FuturesAI Scoreboard: Do UP Calls Beat DOWN Calls?
The page states that it publishes both directions so the record cannot be framed one-sidedly. That works in the reader’s favor here, because it lets you see the disagreement.
What would a fair filter test look like?
A fair filter test writes the rule down before looking at results, requires a bucket large enough that its rough 95% range is narrow, and scores the rule only on calls made after it was written. The scoreboard lists 13 pending calls as of November 18, 2026, which is the kind of new data a rule can be tested against.
- Write one rule in a sentence. For example, “DOWN calls on Korean stocks.”
- Count its calls so far. Under 30, call it a note and not a rule.
- Cross-check it against market and window so it is not a disguise.
- Subtract trading costs. A hit rate means little until it clears the fee line; see FuturesAI Payback: What Fees Take From Margin at 10x.
- Score it only on new calls, and keep the old record out of the test.
Your next step: pick one slice from the table, write its rule in one line, and score it on new calls before you believe it.
See the FuturesAI payback page
FAQ
How many calls does a bucket need before I trust its hit rate?
There is no magic number, but the rough 95% range is a fair guide. A bucket of 23 calls with 6 hits runs from about 13% to 46%, which is too wide to act on. Waiting for a bucket of 100 or more narrows that range, though it never reaches certainty.
Is a filter that worked on past calls likely to work on new ones?
Not reliably. A filter chosen after looking at the results has already been fitted to them. The fair test is to write the rule down first and score it only on calls made afterward.
Does this mean the FuturesAI scoreboard is wrong?
No. These are my own counts from the public ledger and they say nothing about whether the scoreboard is accurate. They show that small slices of a small sample move around a lot, which is a limit on what any reader can conclude.
Sources
- FuturesAI AI Prediction Scoreboard: ledger of 100 records, direction summary and pending queue, read November 18, 2026
- FuturesAI payback page: partner exchange fee rates, read November 18, 2026
Fact-checked based on public sources as of November 18, 2026. Counts are my own tallies of the public ledger and will change as new calls are scored. Educational content, not investment advice.
