Blog
How to fact-check a financial prediction, step by step
By Callsight · Published Oct 9, 2026 · Updated Oct 9, 2026
Figures as of Nov 4, 2026. They change as new windows settle.
As of Nov 4, 2026, Callsight has a verdict on 766 calls by the finance YouTubers it tracks, each reached the same way: find the exact words and the second they were said, take the first close after the video went public, judge it at fixed windows, and compare the result with how often that move happened anyway.
How do you fact-check a financial prediction?
A prediction made on YouTube can be checked like any other statement, once it is pinned down: what exactly was said, when, about what, and by when. The steps below are the ones Callsight applies to every call it records, the same way for every channel. Each is shown on a real call from the record, and the thresholds come from the same source as the methodology page, so the two always agree.
- Find the exact words and the second they were said.
- Pin the moment the video went public.
- Decide what the words call: instrument, direction, target and deadline. Set aside what is too vague to check.
- Take the first close after the video went public as the starting price.
- Check the result at fixed windows, and at the stated deadline if there is one.
- Decide what counts as Correct, Incorrect and Partly correct before you look.
- Compare with the base rate: how often the move happened anyway.
- Count a repeated call once.
- Keep a forecast read out from someone else apart from the speaker’s own.
- Look up the whole record: every call, counted by the same rule.
Step one: find the exact words, and the second they were said
Start from the words as they were spoken, in the video itself. A summary or a clip can change them, and one that turns “could” into “will”, or drops an “if”, is a different prediction, maybe one nobody made. Find the moment and note the time in the video.
The card below is how Callsight records a call. The quotation is checked character by character against the video’s captions, and a call whose words cannot be found there is never published. The play button opens the video at the second the words begin, so anyone can hear them in context. The quotation is never translated, because a translation is a paraphrase.
“I'm raising my price Target to now 260 I think this company is going to go up to 260 this year”Entry price (first close after the video)
$218.46
Jan 13, 2025
Price at close
$247.38
Jan 9, 2026
Return
+13.2%
AMZN was $218.46 on Jan 13, 2025. The target was $260.00. Between then and Jan 9, 2026 the high was $258.60, so the level was not reached.
30 days
Incorrect
90 days
Incorrect
180 days
Incorrect
1 year
Partly correct
The steps below come back to this call.
Step two: pin the moment the video went public
Everything after this is measured from the moment the words became public: the starting price and every window. Use the time the video was published. The call above went public at Jan 10, 2025, 21:22 UTC.
Callsight takes that moment from YouTube’s own record of the upload, and a premiere counts from when it actually started. A call made during a livestream is shown but left out of every rate, because the moment it was said, and so its first price, cannot be pinned down. A video taken down later does not take its calls with it: a claim whose source goes offline keeps counting, since a deleted video says nothing about whether the forecast came true.
Step three: decide what the words call, and set aside the vague ones
A checkable prediction names an instrument (a share, a coin, a fund or an economic figure) and says which way it will go, or what level it will reach. A target and a deadline make it sharper, but neither is needed: the checker can fix the time (step five). Write down what the words say, and nothing they do not.
The AMZN call in step one has all four parts: an instrument, a direction, a target of $260.00 and a deadline, which its card dates to Dec 31, 2025.
Most calls are thinner. The one below gives a direction for BTC, with no target and no date. The record reads it as a call for a higher price and checks it at the fixed windows, like every call.
“it is my belief that the release of this etf marks a near bottom for bitcoin's price”Entry price (first close after the video)
$20,743
Jun 27, 2022
Price at close
$30,692
Jun 27, 2023
Return
+48.0%
S&P 500 +14.1%
BTC closed at $20,743 on Jun 27, 2022. On Jun 27, 2023 it closed at $30,692, a return of +48.0%. SPY returned +14.1% over the same window. The call was that it would rise.
30 days
Correct
90 days
Incorrect
180 days
Incorrect
1 year
Correct
No deadline was stated; the headline is the one-year window.
The instrument can also be an economic figure. The call below says there will be no housing crash. It is checked against the national house price index on FRED (CSUSHPINSA), as first published, where a crash is a fall of at least 10% from its highest figure since the call. Its words name no date, so it too is checked at the fixed windows.
“we're not going to be seeing a 10 to 20% real estate collapse anytime soon outside of a few localized markets.”Reading when said
331.51
Jun 9, 2025
Reading at close
336.66
Jun 9, 2026
Change
+5.15
The house price index (CSUSHPINSA, first releases) stood at 331.5 in June 2025, the figure at the start of the window. Between then and June 2026, its deepest fall from a running peak was −1.5%, in January 2026. A fall of at least 10.0% counts as a housing crash, and at least 5.0% as partial. The call was that no housing crash would come.
30 days
Correct
90 days
Correct
180 days
Correct
1 year
Correct
No deadline was stated; the headline is the one-year window.
Then there is everything that cannot turn out wrong, such as a coin with a bright future or a share that looks cheap. Remarks like these name no direction, or nothing a price could settle. Callsight records and counts them but never scores them, so a vague remark is never marked Incorrect. They are a large part of what gets said. On every channel Callsight tracks, more than half of the statements made in the channel’s own words were too vague to check: from 53.4% of 506 for Graham Stephan to 61.3% of 302 for Andrei Jikh.
Step four: verify a market prediction from the first close after the video
To verify a market prediction, start from a price the speaker could not yet see: the first closing price strictly after the video went public. A close that had already printed when the video appeared is a price the speaker could see, and judging the call from it scores it against the past. The price on the day you watched the video is no better.
For the AMZN call, the first close after the video came three days later, on Jan 13, 2025: $218.46. That is the entry price on its card, and every window is measured from it.
Prices are adjusted for share splits, so a target said before a split is compared like for like. A call on the direction of a share’s price is judged on its total return, dividends included; a coin pays none, so its return is the price change alone.
Step five: check at fixed windows, and at the stated deadline if there is one
Decide how long to wait before you look, and look at more than one length of time. Callsight checks every call at the same six windows, from 30 days to 5 years, whatever the speaker said, and publishes all of them. The headline is 1 year, fixed in advance for every call. A deadline the words name is shown too, on its own line, but it never enters a rate, because a window chosen after the fact, by the speaker or by the checker, can always be chosen to flatter.
Most calls name no deadline at all. Of the 1,044 calls Callsight counts, 135 have a deadline in the words that the card shows; the rest are judged at the fixed windows alone. The windows often disagree. Of the 590 counted calls that are Correct or Incorrect at both 30 days and 1 year, 190 have a different verdict at the two, counted call by call.
A verdict can change more than once, too. The BTC call in step three is Correct at 30 days, Incorrect at 90 days, Incorrect at 180 days and Correct at 1 year.
The AMZN call from step one is Incorrect at the deadline its card shows, Dec 31, 2025, and Partly correct at one year. Both verdicts are shown, and only the one-year verdict enters a rate.
Some targets are reached only after a year. The BTC target below is Incorrect at the deadline its card shows and at one year, and Correct at 3 years. The card shows the windows up to a year; the call’s page shows the rest.
“I think Bitcoin will hit $100,000 just like I said last time”Entry price (first close after the video)
$46,446
Jan 3, 2022
Price at close
$16,675
Jan 3, 2023
Return
−64.1%
BTC was $46,446 on Jan 3, 2022. The target was $100,000. Between then and Jan 3, 2023 the high was $48,190, so the level was not reached.
30 days
Incorrect
90 days
Incorrect
180 days
Incorrect
1 year
Incorrect
In the one-year window the highest price was $48,190, against a target of $100,000.
Step six: evaluate a market prediction as Correct, Incorrect or Partly correct
Write down what would count as Correct before you look at the price. Callsight’s definitions are the same for every channel:
- A call that a price will rise or fall is Correct when the return over the window goes the way it said, and Incorrect when it goes the other way. A move smaller than 0.5% either way is Within band, and counts as neither.
- A price target is Correct when the day’s high reaches the level inside the window (the day’s low, for a target below the price). It is Partly correct when the window’s last close had covered at least 50% of the way from the entry price to the level, and Incorrect otherwise.
- A crash call needs a fall of at least the size its subject sets, from the highest point since the call: for a housing crash, 10% on the house price index. Half that fall is Partly correct. A call that no crash will come is judged the other way round.
- A window that has not closed yet is Pending.
Only Correct and Incorrect enter a hit rate. Partly correct and Within band are shown on each call, and neither counts as a hit or a miss. The AMZN target from step one is the middle case. In its one-year window the highest price was $258.60, short of $260.00, so the level was not reached. But the window’s last close, $247.38, had covered at least 50% of the way from the entry price, so at one year the call is Partly correct. At the deadline its card shows, the last close had covered less than that, and the call is Incorrect there. How often targets are reached, by channel and by how far away they were set, is in Callsight’s article on price targets.
Step seven: why investment prediction accuracy needs a base rate
A Correct verdict says less than it seems when the move was likely anyway. The base rate puts a number on that: how often the called move happened anyway, over past windows of the same length, using only what was known on the day of the call.
The BTC call in step three is Correct at one year. Its base rate at one year is 67.9%: before the day it was made, the price had risen over 67.9% of the one-year spans in its own history, leaving out spans that barely moved. A call for a rise started with better than even odds, whoever made it.
So set a hit rate beside the base rate of the same calls, and give the number of calls. A share of calls that came true means little until it sits next to how often the same moves happened anyway, and one taken over a handful of calls can swing either way. Callsight’s record page prints each channel’s hit rate at every window, with the base rate of the same calls beside it and the number of calls behind both. The comparison across the tracked channels, against base rates and benchmarks, is in the article on whether finance YouTubers beat the market.
Step eight: count a repeated call once
People repeat the calls they believe in, and often the ones that are going well. Counted every time, a call made in ten videos weighs ten times as much as one made once, and a record can be built by volume. Callsight counts the same call on the same instrument in the same direction, inside one fixed block of 90 days, as one forecast however often it was made. Two opposite calls in one block stay two, so a change of mind stays visible.
The call below is one of 16 counted calls from Coin Bureau for a rise in BTC, made between Jan 20, 2024 and Mar 13, 2024, all inside one block. In the channel’s hit rate the 16 count as one forecast.
“even though these allocations are likely to be small in percentage terms in dollar terms they could be enough to perpetually push up btc's price”Entry price (first close after the video)
$66,074
Mar 6, 2024
Price at close
$89,932
Mar 6, 2025
Return
+36.1%
S&P 500 +13.8%
BTC closed at $66,074 on Mar 6, 2024. On Mar 6, 2025 it closed at $89,932, a return of +36.1%. SPY returned +13.8% over the same window. The call was that it would rise.
30 days
Correct
90 days
Correct
180 days
Incorrect
1 year
Correct
No deadline was stated; the headline is the one-year window.
Across the record, 656 counted calls are Correct or Incorrect at one year, taken call by call. With repeats counted once, the channels’ records hold 402 decided forecasts at that window, and each channel’s hit rate is taken over those.
Step nine: keep a bank’s forecast apart from the speaker’s own
Presenters often read out someone else’s forecast, such as a bank’s year-end target. The presenter did say those words, but the forecast belongs to the bank. Crediting it to the person reading it out, or holding it against them, gets the author wrong.
The card below is one of those: Graham Stephan reading out a target set by Goldman Sachs. The card names whose forecast it is and gives no verdict.
“Goldman Sachs just raised their year-end S&P 500 target to 8,000, up from 7,600, which implies solid upside from where we are today.”The record lists 448 forecasts that the channels read out from someone else. Each is shown with its source where the words name one, and none is counted in any channel’s figures. Another 470 are in the channels’ own words and left out of the figures too. Most of them are calls whose direction could not be read with confidence, or whose reading is being checked again; the others include positions the speaker already held, calls made on a livestream and sponsored segments.
Step ten: look up a finance influencer’s track record on Callsight
One remembered hit says little: hits get repeated and misses fade. A track record means every call, counted by the same rule. On Callsight:
- The record page gives each channel’s hit rate at every window, beside the base rate of the same calls and the number of calls.
- Each channel’s page lists its calls with the verdict at every window: Andrei Jikh, Coin Bureau, Graham Stephan and Joseph Carlson.
- Each call has a page of its own with the words and the sentences around them, the verdict at every window with the prices behind it, and a form to report a misreading, such as the page of the call in step one.
- Each instrument and each economic subject has a page with every channel’s calls on it, such as Bitcoin and a housing crash.
If you think a call was misread, the corrections page says how to report it, and anyone whose words are measured can ask for a reply to be shown beside the result.
Can you judge the quality of financial advice, or vet an investment recommendation, this way?
Only the forecast inside a piece of advice can be checked this way: an instrument, a direction or a level, and a time. Advice on how much to save or how much risk to take makes no forecast, and no price can show it right or wrong. Neither can a recommendation to buy that says nothing about what the price will do.
A checked record describes the past. It says how often a channel’s calls came true, beside the base rate and the number of calls, and nothing about whether the next one will. Callsight does not rate people or their advice, and nothing on the site is investment advice: it measures what was said against what prices and published figures did next.
How these examples were chosen
The cards were chosen on Oct 9, 2026, by a rule set down before choosing. Four calls had been proposed for this article before it was written; each was checked again that day and was still listed, the speaker’s own and counted (the relayed one still stored as someone else’s forecast), with no open question about how it was read, so all four are shown. Two steps needed a call those four did not give. For a call with every part, the rule took the most-viewed counted price target with a deadline its card shows and a Partly correct verdict at one year. For a repeated call, it took the largest block of repeats whose one-year windows had all closed, and from it the most-viewed counted call whose words name the price; a call from the same video whose words did not was passed over. Every call is listed on its channel’s page.
Figures are computed by code from the same record as the rest of the site. This article was drafted with the help of AI tools and reviewed by a person before publication. It is not investment advice.