How this is measured

Every figure on this site comes from the same process, applied to every channel in the same way, with no human deciding which videos count. This page describes every rule that decides a verdict, in enough detail that you can disagree with it — and, where it matters, check it.

1. Which videos

We take a channel’s entire upload history, not a selection. Nobody here chooses which videos are examined, because choosing videos is choosing the answer: the memorable ones are the ones somebody got loudly right or loudly wrong, and a record built from memorable moments measures something other than accuracy.

These rules remove videos, and they are the only ones:

• Shorter than 180 seconds, or fewer than 300 words once transcribed: too brief to carry an instrument, a direction and a deadline. A clip under 90 seconds is not fetched at all.
• Published before the channel’s backfill window opens.
• A guest episode. Automatic captions do not say who is speaking, so a video is held back before anything in it is read when its title matches a guest pattern, when the opening of its description names a guest, or when the channel is one whose host has guests on — then every video of it waits. A person looks at each one held and releases it or excludes it as a guest episode; a reviewer can also exclude one the rules missed.
• Every video of a channel made of interviews. Such a channel can be followed, but nothing in it is read, shown under the host’s name or scored.
• No English caption track: there is nothing to quote from.

A guest can still speak in a video that passes these rules. Where we find that words recorded under a channel were a guest’s, or we cannot confirm who said them, the claim is taken off the lists and counts in no record (section 11). If you see words attributed to the wrong person, report it from the claim’s page.

2. What counts as a claim

A claim is a statement that a named instrument or an economic figure will rise, fall or reach a level, or that an event such as a recession or a crash will or will not happen. Everything else is counted and never scored. That last part is a figure we publish rather than hide: the vague share is the proportion of what somebody says that cannot be right or wrong, and on some channels it is the most informative number on the page.

Claims are also left out of every rate when the speaker was reading a sponsor segment, repeating somebody else’s forecast, or describing a position they had already taken; when the call was said on a livestream, where the moment it was said, and so its first price, is uncertain; and while its direction could not be determined with confidence or how it was recorded is being checked again (section 11). A relayed forecast is still shown, with whose it was, because a channel that spends half an hour on a bank’s targets did say those words — it is simply not scored against the person reading them out.

Some statements sound like forecasts and are not: odds (“a 35% chance”), what the market is pricing in, a wish, something already done, a historical average, a valuation. When a review finds one among the claims, it is taken off the lists as recorded in error (section 11). A figure given as odds is never read as the size of a fall a crash call needs, and when a new claim is recorded, a target the words give as odds is dropped. The heading of a claim and its prediction line print a figure only when a rule tested that figure: a price target that names no direction prints none, and a level on an economic figure prints only on the verdicts of the rule that tests levels. Otherwise they give the direction alone, and in some cases not even that: a target the price or the figure had already passed when it was said reads as “around” that figure; a call held out of the score while it is checked again reads as a call under review, with no subject either; a call whose direction could not be determined reads as a call with no clear direction, with no subject either; a call its topic’s rule refused for its subject says that it was not judged and what the rule reads; and words a review found to describe the present are called a remark about the present.

Since 30 September 2026 a machine also looks for those shapes when a claim is recorded, and before an older one can be published: a past purchase, a valuation, a description of the present, a what-if, words just before the quote that play a clip or hand over to another voice, and a recorded reading the words may not give — a topic or an instrument they never name, a target they do not contain, a figure “staying high” recorded as a rise. Such a claim is held out of every record, under a neutral heading, until a person has read it; one already counted keeps counting and is listed for a person to read. Two readings of the same words in opposite directions are both held. That check never decides that something was not a forecast: a person does.

An instrument is read from the words as they were said. A word people say that is also a ticker (“it”, “all”, “key”) is never filed under that ticker, and two or three letters that are both a coin’s ticker and a company’s are filed only once the quote has been read with both in front of the reader; until then the claim is not published.

Some words name a group, not an instrument, and a call that uses one is judged on a stand-in chosen once, from a fixed table, for every channel alike (with or without a leading “the”): “crypto” and “crypto market” on Bitcoin (BTC); “altcoins” on Ether (ETH); “market”, “markets”, “stock market” and “index funds” on the S&P 500 fund (SPY); “tech”, “tech stocks” and “big tech” on the Nasdaq-100 fund (QQQ); “small caps” on the Russell 2000 fund (IWM). A verdict on a stand-in is a verdict on that instrument, not on everything the word could mean, and the card of such a call says which word was recorded and which instrument it was judged on.

3. Saying it twice

The same call about the same instrument in the same direction, inside one 90-day period, is one forecast however many times it is made. A channel that mentions its favourite stock in forty videos has taken one position, not forty, and counting it forty times would let volume buy a record. The periods are fixed blocks of 90 days, so their boundaries do not move with the calls.

This is not a small adjustment. On this site’s first channel it moved the rate down, not up, because a person naturally repeats the calls that are going well. The direction is in the rule, so two opposite calls on the same instrument in the same period stay two forecasts: changing your mind is visible, and it should be.

4. The quotation

How a claim is found. A language model reads each video’s captions and proposes the predictions in it and how each is to be read, including: the words; whether they make a call that can be checked or only a vague remark; what they are about; the direction or level they name; any deadline they state; whether the forecast is the speaker’s own or someone else’s read out; and whether it was said in a sponsor segment or about a position already held. A proposal is only a candidate. The quotation printed on a claim’s page is cut from the caption track itself, never copied from the model, under the rule below. Every claim of a kind that is scored then has its direction, and its instrument unless the site’s own list of names already settles it, read again by a model, with independently worded questions: a change needs two readings that agree, and a claim they cannot settle is held out of every rate (section 11). Whether a call came true is decided by code alone, from published prices and economic data, under the rules on this page; no model is asked. A model can still misread a sentence, which is what the report form on every claim page is for.

Every claim carries the speaker’s own words and a link to the second they were said. The quotation is checked, character by character, against the caption track: a paraphrase is refused, and a claim whose quote cannot be located is discarded rather than published. Quotes run between 15 and 400 characters.

Quotations are never translated, in any language edition of this site. A machine translation is a paraphrase with a hundred per cent substitution rate, and putting one inside quotation marks beside a real person’s name would be the exact misattribution this process exists to prevent.

5. Prices

A forecast is judged from the first closing price strictly after the video was public — never the same close if that close had already printed. Using a price the speaker could already see would be scoring them against the past, and it is the first thing anyone disputes. An instrument that was not yet trading when the call was made has no such close, and the call is Not verifiable; so is a price target whose first close after the video comes more than 5 days late, after a hole in the data.

Every closing price printed on this site is adjusted for share splits, so the verdict and the chart show the same price for the same day. A target is printed as the speaker said it, and judged after restating it for any later split, so a target spoken before a split is compared with the same target after it. The return on a call about a price’s direction is a total return, with dividends included, and is labelled as one (a coin pays none, so a coin’s return is a plain one and is labelled a return); a price target is judged on prices alone.

6. Windows

Every claim is checked at the same fixed windows, whatever horizon the speaker mentioned: 30 days, 90 days, 180 days, 1 year, 3 years and 5 years. The longer windows are judged once a claim is old enough to have reached them. All of them are published. Reporting one window would mean choosing one, and the choice would be invisible — a call wrong at three months and right at a year is a different result from one right at both. The headline figures and the default ranking read the result at 1 year.

A window is counted in calendar days from the day the video was published, and it ends on the last close on or before that day. A price call’s return is measured from the first close after the video went public (section 5), which for a video published after the close or at a weekend comes a few days later, so the window of such a call runs that many days fewer than its length. The count is the same for every call, and the chart on a claim’s page shades exactly the days that were judged.

When the words name a deadline — a date or a period — the claim’s page also shows a window ending there, as the stated deadline. It never enters a rate or a score, because a window the speaker chose would let anyone pick the one that flatters them. Otherwise the page says the call is checked at the fixed windows and prints no date.

A window that has not closed is Pending. Until a call’s result at 1 year is in, its headline verdict comes from the longest shorter window that has one, and says “so far”. A channel whose oldest call is younger than a window has no rate there, rather than a rate of zero: the page shows the date its first result can exist. The base rate beside a long window rests on fewer independent windows of history, and the count behind it is stored with it.

Still correct after 1 year. For every call on the direction of a price that was Correct at 30 days and has a settled 1 year result, we count whether that later result was also Correct, and print the share with its sample size, beside a base rate for the same instruments: how often, once an instrument had moved one way over 30 days, it was still that way after 1 year. The opposite movement — Incorrect at 30 days, Correct at 1 year — is counted beside it. Price-target calls are left out: a target reached inside 30 days is still reached inside 1 year, so they could never move. With fewer than 20 such calls, a count is printed instead of a percentage. Repeats count once, as everywhere on this page, and closing prices are used throughout. The figure is context and enters no score or ranking.

7. How a call is judged

Each window gets one verdict. Correct and Incorrect are the only ones that enter a hit rate. Within band means the move was too small to call either way: a figure that barely moved has not confirmed a call, and it has not refuted one. Partly correct is real partial credit: a price that covered at least 50% of the distance to its target without reaching it, or a fall of at least half the size a crash call needed. Both are shown, and neither is counted as a hit or a miss. Pending means the window has not closed, or its figures are not yet published.

8. When a call cannot be checked

Not verifiable means no rule here can measure what was said. The claim stays on its page with the reason, and it counts in no rate. It is preferred to a guess: a missed verdict can be added later, a wrong one beside a real name cannot be taken back. The reasons:

• No instrument or no direction. A call that names no instrument, or does not say which way it goes. An interest-rate call that does not say whether the rate goes up or down is not read as a call for a hold.
• A comparison. “X will beat Y” is not checked yet, and it is never judged as a plain rise of X.
• A target that is not a price of the instrument. A percentage, a multiple, an index level, or a commodity price quoted against a fund (gold per ounce is not a share of a gold fund). So is a target more than 5 times outside the instrument’s own price range, and, for an economic figure, a target in a unit its series is not quoted in, or more than 5 times the largest figure the series had shown.
• A target already met when it was said, on the side the call named, for a price or an economic figure: the window could not test it.
• Not what the rule measures. Economic rules read US national figures and the crash rule reads the US stock market in general. So a call whose words name another country, central bank or currency, or one city, state or local market, is not judged on US figures, even when it is also about the US; nor is a call about bond, Treasury or mortgage yields filed as interest rates in general; nor, on the crash rule, a call whose words name crypto, or whose recorded subject is one sector or another market. A missed verdict is preferred to a wrong one. These are found by fixed word lists, and by a review whose findings can only make a claim Not verifiable, never Correct or Incorrect; each such claim’s page gives the reason, and section 15 counts the review’s findings by category.
• Missing prices. No prices for the instrument, an instrument not yet trading when the call was made, or a window whose last price comes more than 5 days before its end, with prices missing in between.
• No rule for this kind of claim, such as a company event, or a statement that could not be wrong as it was said, such as one about the present.

Expired is kept for a window that outlives the instrument it is about, and the current rules do not use it: such a window closes at the last price when the company was bought for cash, at $0 after a bankruptcy or a coin’s collapse, and is otherwise Not verifiable.

9. The base rate

A hit rate on its own says little, because some moves happen far more often than others, whoever calls them. So every rate here is published beside the base rate: how often the called move happened anyway in comparable past windows, measured over the same window lengths, using only what was already known on the day the claim was made. Past windows whose move stayed within the band are left out of it, as they are left out of a hit rate, and no base rate is published from fewer than 50 past windows.

• The economy. How often the series moved that way over a window of that length — inflation, unemployment, mortgage rates, house prices — or, for the Fed, how often the rate was actually hiked or cut, and how often a recession was dated. Measured from the published series on FRED back to 1980 (the Fed’s target range from December 2008, when it began), at the values first released rather than at the figures later revised.
• A fund, a coin or a single company. How often that fund, coin or company’s shares rose over a window of that length, from its own daily history. This is deliberately not the same question as “what share of shares rose”: nobody buying an index fund held a basket of individual bets, and the fund’s own record is what they actually faced. Every instrument anyone here has made a call about gets its base rate from the night after its first call, measured over its whole price history before each call. A call made when the instrument had fewer than 500 daily closes has none, however long it has traded since. Where a stored price series has a break that may be two instruments under one name, no window across the break is counted, and a call after it is measured on the prices after it alone.
• A target. Measured against how often the instrument or the series moved the called way at all. Reaching a level is rarer than moving towards it, so this is a stricter bar for the speaker, never an easier one.

A company’s base rate is its own past, not its peers’. It answers “how often did this company’s shares move this way over a window this long”, not “how often did a comparable company’s”. Measuring the second honestly needs the prices of every company that was comparable at the time, including the ones that were bought, broken up or wound up — and the free price sources keep only the survivors. Measured: of the S&P 500 as it stood in June 2016, 34 per cent is no longer in the index, and of those we can price 47 out of 242. A company’s own history has no such gap, since a company that fails simply stops having prices. It does carry one bias, stated plainly: the companies people talk about are often the ones that had been rising, and their own past says so, which makes a bullish call on them a harder bar to clear, and a bearish call on them an easier one by the same amount. Before methodology 0.3.1 (section 15) a company call had no base rate at all. A verdict that settled before 30 September 2026 keeps the base rate it was stored with; on coins, a few of those rest on fewer than 500 closes before the call, because the rule then counted the instrument’s whole stored history.

A crash call has no base rate, and the column is blank. Neither does a call on an instrument with too little history. A call without a base rate still counts in the plain hit rate; it cannot be above or below a base rate, so it stays out of the comparison described next.

One limit, stated plainly: the Fed’s base rate is unconditional, and its history starts in December 2008, so it leans on 2009 to 2015, when the rate sat near zero and barely moved. It treats a hike or a cut as rarer than they have been since, and a correct Fed call can count for more than it deserves.

10. The score and the ranking

The score is a summary of how a channel’s checked calls compared with the base rate; it describes the record, not the person. It combines three parts: the hit rate above the base rate (40%), the return against a benchmark (40%) and calibration (20%), each pulled towards no edge at all by how little evidence there is — the hit rate as if every channel started with 20 calls exactly at the base rate. Fifty means no better than the base rate.

The hit rate and the base rate it is compared with are measured on the same calls: the decided calls that have a base rate. Every base rate is printed as resting on m of n calls. When fewer than 50% of a channel’s decided calls have one, the base rate describes too little of the record: the channel has no score and is not ranked, and the page says why: “Not ranked: only m of n calls have a comparable base rate”. A channel whose calls are mostly on instruments with too little history, or on crashes, is in that position.

The return against a benchmark is the instrument’s total return minus the benchmark’s over the same window, counted in the direction called: the S&P 500 for shares, funds, Bitcoin and Ether, and Bitcoin for other coins. Only calls on a price’s direction whose instrument has a benchmark carry one, so every average is printed with the number of calls behind it, and below 5 calls the count is printed and no average.

A channel is ranked only after 20 decided calls, repeats counted once. Below that the rate is still shown, and marked as not yet ranked — hiding a thin record is indistinguishable from having something to hide. A channel counts as having a proven record with a score of at least 60 on at least 20 calls, and a signal needs 2 such channels agreeing on an instrument, with no channel that has a proven record calling it the other way. In the weighted consensus a channel’s call counts in proportion to how far its score is above fifty, and not at all at fifty or below or under 20 calls; its weight also halves with every 30 days of the call’s age, and is a little higher for a call put with certainty than for a hedged one. The proven status and those weights use a channel’s whole record, economic, stock and crypto calls together, so a channel can reach them on the strength of its economic calls and then weigh in on a coin.

11. Corrections

A settled verdict is frozen. It changes only through a published correction or a versioned re-run of the rules, and both leave a record: every verdict a re-run rewrites keeps its earlier version, and section 15 lists what the last one changed. There is no path — and no database column — that would allow a score to be edited quietly.

A claim is taken off the lists when its source is taken down at the request of whoever holds the rights, or when we find we recorded it in error. Its page stays at the same address, with the date and the reason, so anything citing it still leads somewhere true. A claim taken down with its source continues to count towards the record: a source going offline says nothing about whether the forecast was right, and a record that improves when its subject deletes the evidence is not a record. A claim we recorded in error — a guest’s words, a speaker we could not confirm, a statement that was not a forecast, a second reading of a sentence already counted — counts in no record. A claim being checked again, because the words may not say what was recorded: the direction, the subject or a figure, stays on its page and is held out of the score until the check is done. So is a claim whose direction could not be determined with confidence: its page says so, and it counts in no rate. So is a claim a machine held because its words look like something other than the speaker’s own forecast, or read two ways (section 2): it waits for a person, who puts it back as it was recorded or corrects it through these steps.

12. Right of reply

Anyone whose words are measured here can dispute any claim: a misattribution, a quote taken out of context, a deadline read wrongly. Every claim page has a report form, and the corrections page says what happens to a report. When a report shows we misread a claim, the claim is corrected through the steps in section 11 and its page says why. A creator who asks us to publish a response will have it shown beside the measurement, not buried on another page.

13. Independence

The ranking is free, permanently, including the top of it. Nothing about a channel’s position can be bought, and no channel measured here is a customer or a partner. If that ever changes it will be stated on this page before it takes effect.

14. What this is not

This is not investment advice. It is a measurement of public statements. Nothing here is a recommendation to buy or sell anything, and a high score is a description of somebody’s past record, not a prediction about their next one.

15. What changed

What changed in 0.3.0

Methodology 0.3.0 was applied on Sep 28, 2026, in two forced re-runs: every verdict was first judged again on corrected data under the old rules, then under verdict rules version 4. The changes:

• Prices were reloaded from one source with dividends restated, claims filed under the wrong instrument were moved to the right one, repeated copies of three economic series were removed, and every verdict was judged again on the corrected data under the same rules.
• Every call is checked at the fixed windows, from 30 days to 5 years, and only those count in a record. A window at the call's own deadline is shown only for calls recorded with a deadline, and it never counts in a record.
• A call that an economic figure will reach a level is checked on the side the call names: a rise must reach the level from below, a fall from above. A level with no side, or in a unit or size the series cannot show, is not verifiable.
• Only a dollar price target is judged as a price. A percentage, a multiple, an index level or a commodity price quoted against a fund is not verifiable, and so is a target the price had already passed when the call was made.
• An economic call about another country or about one local market is not verifiable against US national figures. A call about bond, Treasury or mortgage yields that was filed under interest rates in general, rather than under mortgage rates, is no longer judged as a Fed decision; it is not verifiable. Claims the review found are marked not verifiable too, each with its category; the review cannot mark a claim anything else.
• Verdict sentences print split-adjusted closing prices, the basis of the chart, and call the return what it is: a total return with dividends.
• A call that one asset will beat another is not verifiable until comparisons are checked; it is no longer judged as a plain rise or fall.
• An interest-rate call that does not say whether the rate goes up or down is not verifiable; it used to be read as a call for a hold.
• A window whose last price comes more than five days before its end, with prices missing in between, is not verifiable.
• A crash or correction call is judged on the S&P 500 only when it is about the US stock market in general; a call about crypto or another market is not verifiable. When the speaker named the size of the fall, and it is larger than the standard one and not a probability, that size is required.
• Interest-rate moves in verdict sentences are printed to two decimals and rounded as the site rounds them.
• An economic target the figure had already reached, on the side the call named, when the call was made is not verifiable.
• The score compares the hit rate with the base rate on the same calls, those that have a base rate, and a channel where fewer than half of the decided calls have one is not ranked.
• The average excess return is shown with the number of price calls it rests on.
• The share of vague statements, and the counts of pending and not verifiable calls, cover only statements on the record.
• One rule decides which channels are credible; the fallback rule was removed and the signal check uses the same rule.

What that did to each channel’s published verdicts. The counts cover the calls a record is built from: the channel’s own calls of the kinds that are scored, at every window a reader is shown, never a relayed forecast, a vague remark or a call that is not scored. Each window counts once, from its result before the two re-runs to its result after them, and a window that ended where it started counts nowhere. “Now Incorrect”, “Now Correct” and “Now not verifiable” count published verdicts by their new outcome; “Other” counts verdicts that are now Partly correct or Within band, or open again; the last column counts windows that were still open and are now Not verifiable.

And, per channel, the claims a review marked Not verifiable, by category, and the claims taken off the lists or held out of the score. Each claim is counted once: one taken off or held is counted there, and not among those reviewed as not verifiable.

• Andrei Jikh — reviewed as not verifiable: 2 (bond, mortgage or savings rates 2). Taken off as recorded in error: 11 (a guest's words 1; speaker not confirmed 1; not a prediction 9). Held out of the score while they are checked again: 7.
• Coin Bureau — reviewed as not verifiable: 14 (crypto under the stock-market crash rule 4; a local market 1; another country 3; bond, mortgage or savings rates 6). Taken off as recorded in error: 18 (a guest's words 2; not a prediction 15; a second reading of a sentence 1). Held out of the score while they are checked again: 42.
• Graham Stephan — reviewed as not verifiable: 17 (a local market 10; bond, mortgage or savings rates 7). Taken off as recorded in error: 44 (speaker not confirmed 5; not a prediction 30; a second reading of a sentence 9). Held out of the score while they are checked again: 27.
• Joseph Carlson — reviewed as not verifiable: 0. Taken off as recorded in error: 26 (speaker not confirmed 2; not a prediction 24). Held out of the score while they are checked again: 11.

Methodology version 0.3.2, verdict rules version 4. Every threshold on this page is read from the files the pipeline runs on, or checked against its code by a test, so the page and the rules cannot quietly disagree.