“80% accuracy”: what a crash predictor is worth against chance

What exactly do sellers promise?

Since 26 October 2026, we have kept a register of public claims by sellers of predictors and signals for crash games: each claim is logged in the seller's own words, with its date, the game it targets, the cash-out target where one is stated, and the advertised price.⁠^[1] As of 30 October, the register holds 19 rows. The sellers' identities stay in our internal register — the full list is available on request — because this study scores claims, not people: what follows keeps only the rates, the targets, the prices and the mechanics.

Four rows advertise a numeric accuracy rate. Only one of them also states its target: “80%” accuracy, for a cash-out at 1.20x on Aviator. The other three announce 90 to 97%, 91% and more, and a 99% “win rate”, with no cash-out target at all — at advertised prices of $500, $1,500 and $5,500.⁠^[1] The remaining 15 rows put no number on accuracy: they sell a language of method (“pattern detection across 50+ variables”), tiers of access to signal channels, promises of earnings, or — in one row logged on 30 October — a staged demonstration in which a second phone displays the crash in advance.

The rest of the register gives the market's scale: subscriptions from $9.99 to $199 a month across the logged offers, free signal channels that funnel readers toward paid versions, and applications distributed outside the official stores.⁠^[1] The scoring turns on the four numeric rates — starting with the question the sales pages never ask: what score do you get by buying nothing?

What score do you get with no tool at all?

Aviator, the game the four numeric claims target, publishes a payout rate of 97%.⁠^[2] The distribution law of crash games is simple, and it is the core of this study: the probability that a round reaches at least the multiplier x is the payout rate divided by x. A player who always cashes out at the same target — no signal, no application, no information of any kind — therefore succeeds in r ÷ x of rounds. That is the no-tool baseline, the “free chance” every predictor must be compared against, target by target.⁠^[3]

Two perimeter notes, because they change how the table reads. First, the second column is computed with the published payout of the game the claims target (97%); our deposited corpus of one million rounds, built on a 98% model, is used here only to check the law itself: its empirical rates depart from the r ÷ x formula by at most 0.20 points across the table's six targets.⁠^[3] Second, the third column is the bound for a game with no house edge at all (1 ÷ target): we always name it the “no-edge bound”, and it is never mixed, inside a single claim, with the 80.8% baseline, which includes the published edge. Our full workings — baselines by target, break-even lines and decisive sample sizes — are attached to the article in the downloadable file.⁠^[4]

The table reads directly: the lower the target, the more often free chance “succeeds”. At 1.20x, a player with no tool wins four rounds in five. That is exactly where the advertised percentages look impressive — and exactly where they must be scored.

Advertised accuracy or useful accuracy: the hit rate does not state the profit

Let us score the one claim in the register that accepts being scored. “80% accuracy” at a 1.20x target, against a no-tool baseline of 80.8% at that same target: the gap is −0.8 points.⁠^[1]⁠^[4] The advertised figure therefore sits below the level reached, free of charge and with no prediction whatsoever, by a player who simply cashes out at the same place every round. And that is only the first of the two comparisons that matter.

The second is the break-even line. Following signals at target x makes money only if the true success rate exceeds 1 ÷ x — at 1.20x, 83.33%. If the advertised 80% rate were exact and durable, every followed round would lose 4.0% of the stake in expectation (0.80 × 1.20 − 1). The free baseline itself loses 3.0% per round: that is the house edge, no more and no less. In other words, a predictor can advertise a high rate, sit below free chance, and lose money faster than the game alone — all three at once.

How many rounds does it take to tell a tool from chance?

Now suppose a seller in good faith, who states a rate and agrees to prove it. How many signals must be tracked before the rate separates from free chance? The answer depends on the gap to be proven: the closer the claimed rate sits to the baseline, the longer the sample must run — the sample size grows with what the target and the gap demand. At a 1.20x target (baseline 80.8%), with a one-sided test at the 5% level and 80% power, our calculations give:⁠^[4]

The other side of this table concerns the usual “proof”: a session screenshot, a week of signals, a hundred filmed rounds. Across 100 signals, the 95% confidence interval around a rate of 80.8% spans ±7.7 points: such a demonstration cannot separate 80% from 88%, let alone from the baseline itself.⁠^[4] Not one of the register's 19 rows declares a sample — no signal count, no period, no measured rate over a complete series, failures included.⁠^[1] Yet the decisive test is easy to state: signals announced before the rounds, tracked over a sample declared in advance, whose rate beats the target's baseline by a margin that survives the size of that sample.

The anatomy of a sale: free signals, a paid bot, the “KYC” deposit trap

The register and the accounts we gathered also describe a sales path in four steps — a pattern that recurs from one offer and one account to the next, without our being able to tie every step to the same seller. We describe it as a mechanism — steps, observed prices, tipping points — naming neither a product nor a seller.

1. The free signals. A public channel posts cash-out signals and screenshots of winning rounds. Entry is free; the step's job is to recruit, and the screenshots never show the complete series.
2. The paid bot. The signals “at higher multipliers” and the full tool sit behind a payment: subscriptions from $9.99 to $199 a month in the register, and up to $5,500 for the offers advertising the highest rates.⁠^[1]
3. The “KYC” deposit. At the moment the buyer tries to withdraw money, a “customer care” contact demands a further deposit in the name of an identity check (“KYC”) before anything can be released. This step rests on user accounts gathered during our monitoring of forums and discussion threads in October and November 2026: qualitative reports, small in number, which we could not verify one by one — we describe it as a reported pattern, not as a measured frequency.
4. The block. In those same accounts, the verification deposit releases nothing: further requests go unanswered, then contact is cut. The money already committed — the subscription and the followed stakes — is joined by the lost deposit.

Each step has its own function. The free signals build trust with chosen images. The payment converts that trust into regular revenue. The “KYC” deposit arrives at the only moment the buyer asks for money back — a second payment, demanded under an administrative pretext, from someone who has already paid twice: for the tool, then for the stakes. And at every step, the previous section's test stays available and free: ask for the target, do the division, demand the declared sample.

Our analysis

Everything above rests on the cited sources and on our calculations. What follows is our analysis, and it mixes three things that we distinguish at every step: the cited sources' findings (the claims register and the game's public documentation), credited to them and cited; our calculations on their published data, flagged as ours, with their margins; and our interpretations, labelled as such.

The register's only scorable rate fails the test it invites. Our calculation, on the register's figures⁠^[1]: the “80%” claim at 1.20x finishes 0.8 points under the free baseline of its own target, and 3.33 points under that target's break-even line. It is not “almost” at the level of chance: it is below it, on both comparisons at once.

Beating free chance is not enough — the line to beat is break-even. Our calculation: a predictor whose true rate were 82% at 1.20x would beat the baseline (80.8%) and still lose roughly 1.6% of the stake on every round (0.82 × 1.20 − 1). Between the baseline and the break-even line stretches a 2.50-point zone where a tool can be “better than chance” and remain a losing purchase on every round. Sales pages never show that zone: they compare their rate to nothing.

The sample size grows with what the proof targets. Our re-cut of the section 4 table⁠^[4]: separating a true 90% from free chance takes roughly 97 signals; a true 85%, roughly 518; a true 83%, roughly 1,978. The most modest-looking rates are therefore the most expensive to prove — and they are exactly the ones short demonstrations claim to establish.

Our reading: a percentage without a target is not a weaker claim than a percentage with one — it is a claim from which the test has been removed. The register's three highest rates (90% and more, up to $5,500) are also the only three that cannot be scored. We do not read that coincidence as an accident of wording: stating a target costs the seller one division, and that division is public.

Our reading: the funnel's last step changes what is being sold. While the product is signals, the buyer pays for information of no measurable value; at the “KYC” deposit, the buyer pays to release his own money — the product is no longer a prediction, but the exit from a blockage the seller controls. It is at that step, not at the advertised rate, that the loss becomes final in the accounts we gathered.

Methodology and limits

The register. Public claims by sellers of predictors and signals, logged between 26 and 30 October 2026 in their own words, with date, game, target where stated and advertised price; no claim without a dated public source enters the register. The sellers' identities stay in the internal register — the full list is available on request — and the published table keeps only rates, targets, prices and mechanics.⁠^[1]

The baseline. For a target x and a game with published payout r, the no-tool success rate is r ÷ x. This study's rates use Aviator's published payout (97%)⁠^[2], the game the numeric claims target. The law itself is checked on our deposited corpus of one million rounds (a 98% model — a distinct perimeter, never used as an Aviator baseline): a maximum departure of 0.20 points between the empirical rates and the formula across six targets.⁠^[3]

Scoring. The scoring compares the advertised claim with the baseline for its target; it does not measure the sold tool's real performance, which no public sample makes observable.

The tests. Break-even: p × x − 1 > 0, i.e. p > 1 ÷ x. Decisive-beating test: a one-sided binomial test, level α = 5%, power 80%, n = ⌈(z_α√(p₀q₀) + z_β√(p₁q₁))² ÷ (p₁ − p₀)²⌉ with z_α = 1.6449 and z_β = 0.8416; p₀ is the target's baseline, p₁ the rate to be separated. The expectation used is gross, per round and before the tool's own price — any subscription paid subtracts further still. The attached file lets any reader redo every number: crash-predictor-claims-scored-2026-11-02.csv.⁠^[4]

Limits. We neither buy nor install the tools sold: the scoring bears on the public claims, not on code we have not run. Three of the four numeric rates cannot be scored for want of a target — the study states that as a result, not as an excuse: a claim without a target has no baseline. The “KYC” deposit step rests on user accounts gathered during our monitoring, qualitative and small in number; we publish it as a pattern reported in those accounts, with that reservation stated in the text itself, and not as a measured share of sales. The baselines hold under these games' documented distribution law (r ÷ x), the law of our deposited corpus: this study is not an audit of the live product, whose internal implementation cannot be observed from outside. Finally, baselines are long-run probabilities: over a short session, chance produces streaks that depart widely from them — which is precisely why section 4 fixes sample sizes.

Gambling involves risks: financial loss and addiction. 18+. If you have already paid a “verification” deposit, keep screenshots of the demands and the payments, make no further payment, and report the facts to your bank — no legitimate release is bought with a new deposit.

Frequently asked questions

References

1. Predictor-claims register kept by us since 26 October 2026: public claims by sellers of crash-game predictors and signals, logged in their own words, with date, game, cash-out target where stated and advertised price (19 rows as of 30 October 2026); the sellers' identities are held in the internal register — the full list is available on request. ↑
2. Spribe — public documentation of the game Aviator: published payout rate of 97% and fairness mechanism (“provably fair”). spribe.co ↑
3. Dagor, A. — Crash-game bust multipliers: a one-million-round simulation, Zenodo, DOI 10.5281/zenodo.23597144 (deposited corpus of one million rounds; a 98% payout model — the distribution law is checked there, while this study's baselines use Aviator's published payout). ↑
4. Our calculations for this study — baselines by target (payout ÷ target), break-even lines (1 ÷ target), gaps for the scored claims and decisive sample sizes; attached file: crash-predictor-claims-scored-2026-11-02.csv. ↑