Methodology

1. What counts as a prediction

From algorithmic prediction sites we record the numeric price forecasts they publish for specific dates (for example a 30-day or end-of-year target), captured by an automated daily snapshot. From YouTube channels and X accountswe extract explicit calls from video transcripts and public posts — a named asset, a stated direction, and where present a concrete price target and time frame. Every extracted call then passes a verification review against the original source: the quoted words must appear verbatim, and sarcasm, questions, conditionals and vague statements (“could go higher”) are rejected. Clear-cut cases are decided by an automated reviewer; anything ambiguous is deferred to human review before it is published.

2. Evidence

Each prediction is stored with its value, the timestamp we captured it, the source URL, and a raw snapshot of the page or transcript taken at capture time. The archive is written before the outcome is known and is never edited afterwards. Every graded row on this site links to its source. For YouTube videos and X posts the prediction timestamp is the platform’s public publication time — the dated recording itself is the evidence — so calls extracted from a channel’s earlier videos are graded from the day the video was published.

3. Grading

A prediction is graded only after its target date has passed, using the daily market price from CoinGecko. Two metrics are computed:

• Absolute percentage error— |predicted − actual| ÷ actual. A forecaster’s headline score is the average of this error across all of their graded predictions (lower is better).
• Direction hit — whether the predicted move (up or down relative to the price on the day the prediction was made) matched the actual move by the target date.

Because short-horizon predictions are structurally easier — a one-day forecast close to the current price is nearly always “accurate” — forecaster pages also break accuracy down by horizon length (1–7 days, 8–30 days, 1–3 months, 3–12 months, over a year), measured from the day the prediction was captured to its target date. Headline averages include all graded predictions regardless of horizon.

4. Exclusions

• Predictions whose target date had already passed when we captured them are never graded — the outcome was already knowable.
• Assets without a reliable market price on the target date (for example delisted coins) are skipped, not counted against anyone.
• YouTube calls without a concrete price target and date appear on a forecaster’s page as verified calls but are not included in error scores.
• Predictions more than 100x away from the asset’s market price on the day of capture — in either direction — are not graded. In practice these are glitched rows on the source page or a ticker shared by two different coins, not real predictions about the asset we price; grading them would corrupt the averages. Genuine long-shot calls stay well inside this bound. The raw pages remain archived either way.

5. Consensus pages

A coin gets a consensus page when at least two tracked sites have published predictions for it within the last 7 days. For each site we show its latest snapshot, matching targets to the horizon nearest to ~1 month, ~3 months, ~1 year (within a tolerance of 15, 45, and 120 days respectively) and to the end of the current year (within 45 days; when fewer than 60 days of the year remain, the column points at the next year’s close). The consensus value is a weighted average: each site’s weight is 1 ÷ (historical average error + 0.05), using its per-coin record when it has at least 3 graded predictions for that coin, its overall record otherwise, and a neutral prior of 30% error when it has no graded history yet. The 0.05 term prevents a single near-perfect record from dominating the average.

6. What we do not do

We publish measured accuracy, not characterizations. A low score means a forecaster’s past public predictions deviated from actual prices by the stated amount — nothing more. Nothing on this site is investment advice, and past accuracy does not predict future accuracy.