I Tried to Score 30 Crypto Forecasts. I Could Defend 5.

I set out to score 30 dated crypto predictions and report the hit rate by claim type. I could not defend 30. I can defend 5 forecasters and 8 versions of their calls. That is too few to say whether law calls beat price calls. It is enough to show something else: the hard part of scoring crypto forecasts is not the outcome. It is the wording of the call.

I changed my thesis for that reason. My working thesis said that fewer than half of 30 calls name a date and a scoring rule, and that law calls score better than price calls. I cannot test the first half with this sample, because I filtered for dated calls. I can only report the second half as a direction, and I give it very little weight.

I read all sources below on 2026-10-11.

Question

Which public crypto forecasts can a stranger score, and how do they score?

I split "can be scored" into three parts, so that the word has a test:

1. Date. The call names a day or a clear deadline.
2. Rule. The call names the event or data that settles it, such as a price on a named venue, or a signed order from a regulator.
3. Probability. The call gives a number between 0 and 1, so I can compute a Brier score. The Brier score is the squared gap between the stated probability and the outcome (1 if it happened, 0 if not). Lower is better. A constant 0.5 scores 0.25.

A call with only a date and a vague rule can still get a hit or miss. It cannot get a calibration score.

Method

Search strategy. I searched the open web for well-known dated calls on two claim types: price calls (bitcoin price at a date) and law calls (a regulator or court acts by a date). I used standard web search, then opened the pages I rely on. I did not use a fixed archive or a random draw. This is the main weakness, and I return to it under Limits.

Inclusion rules.

• The call must carry a date, found in a source I read.
• The outcome must be checkable in a source I read.
• I count each forecaster once, but I list revised versions of a call separately, because a revision is a new claim with its own date.

Exclusion rules. I dropped calls with no date. I did not count those in the sample. That is why the sample cannot tell me how common undated calls are. I found many undated "bitcoin will go much higher" claims in headlines. I did not log them, and I cannot give a share.

Source rule. For each outcome I wanted three independent sources. I met that rule for the 2025 bitcoin price (two news sources, and a third forecaster-side source), and not for the others. I flag each weak spot in the text.

Arithmetic. I computed every Brier score by hand, without the Lab. You can repeat each one with this formula:

$B = (p - o)^2$

Here $p$ is the stated probability and $o$ is 1 if the event happened and 0 if not.

Findings

The ledger

#	Forecaster	Call (date made)	Resolution date	Type	Probability stated	Outcome	Score
1	John McAfee	Bitcoin at $1 million [1]	End of 2020	Price	No	Miss	Not computable
2	Tom Lee	Bitcoin at $25,000 [2]	End of 2018	Price	No	Miss	Not computable
3	Standard Chartered (Kendrick)	Bitcoin at $200,000 [4]	End of 2025	Price	No	Miss	Not computable
4	Standard Chartered (Kendrick)	Same target, cut to $100,000 on 2025-12-09 [4]	End of 2025	Price	No	Miss	Not computable
5	Bloomberg analysts	Spot bitcoin ETF approved [7]	2024-01-10	Law	0.90	Hit	0.0100
6	Bloomberg analysts	Spot ether ETF approved by May [9]	2024-05-23	Law	0.70 (Jan 2024)	Hit	0.0900
7	Bloomberg analysts	Same event, odds cut (Mar 2024) [10]	2024-05-23	Law	0.25 or 0.30	Hit	0.5625 or 0.4900
8	Bloomberg analysts	Same event, odds raised (2024-05-20) [11]	2024-05-23	Law	0.75	Hit	0.0625

Rows 5 to 8 involve two sets of analysts working together at one firm, so they are one forecaster group, not two. The ETF rows are the only ones with probabilities.

Price calls: four versions, four misses

McAfee said in 2017 that bitcoin would reach $500,000 by the end of 2020, then raised the target to $1 million. ForkLog reported on 2021-01-01 that bitcoin had not reached $1 million and traded near $29,500 [1]. That is about 3 percent of the target (29,500 divided by 1,000,000 is 0.0295, my arithmetic). The same page says he stood by the call for a time and then backed away from it. The page does not give a closing price for 2020-12-31, so my figure is for the first day of 2021.

Lee's call is the second price case. CryptoSlate carries his statement that bitcoin could reach $25,000 by the end of 2018 [2]. A Bloomberg headline from 2018-12-13 says he held that "the market is wrong" [3]. I could not open that page (HTTP 403), so I use it only for its headline. Coverage that my search returned put bitcoin near $3,400 in December 2018. I did not verify that number in a primary dataset, so treat it as unconfirmed. Even so, the gap is far larger than any data-feed error.

The Standard Chartered case is the one I find most useful. The bank's analyst Geoffrey Kendrick held a year-end 2025 target of $200,000. On 2025-12-09, 22 days before the deadline, the bank cut it to $100,000 [4]. Decrypt gives the stated reasons: buying by corporate treasuries had "run its course", and quarterly ETF inflows had fallen to 50,000 BTC [4].

Bitcoin Magazine reports that bitcoin entered 2026 near $87,000 [5]. ARY News gives a close of $87,696 [6]. These two sources differ by a small amount and agree on the level. Both versions of the call missed. The revised version missed by about 12 percent: (100,000 minus 87,696) divided by 100,000 is 0.123, my arithmetic.

I note one thing in the revision's favour. It was a better forecast than the original. It was closer, and it came with reasons I can check. A cut 22 days out is still not much of a risk, and I would not give the bank much credit for it. I score it as a miss on a revised claim, not as a recovery.

None of these four calls gave a probability. That means none can earn a Brier score. All I can say is hit or miss, and a miss on a point target is the expected outcome for almost any point target. A forecaster who says "$200,000" and is off by 56 percent has not told me how surprised to be.

Law calls: four versions, four hits, three with a clear rule

The Bloomberg ETF analysts gave a stated probability and a stated date. Cointelegraph quotes the 90 percent odds of approval by 2024-01-10 [7]. The SEC approved spot bitcoin ETFs on 2024-01-10, per a Mintz note [8]. The Brier score is $(0.90 - 1)^2 = 0.0100$.

The ether ETF series is more interesting, because the odds moved:

• January 2024: 70 percent for approval by May [9].
• March 2024: cut. The Block's headline says 30 percent [10]. A search summary of other coverage gave 25 percent. I could not reconcile the two. I score both: 0.5625 at 25 percent and 0.4900 at 30 percent.
• 2024-05-20: raised to 75 percent for approval that week [11].

On 2024-05-23 the SEC staff approved the 19b-4 rule changes for eight ether ETFs, according to Mayer Brown [12]. So the event happened. The three scores are:

$(0.70 - 1)^2 = 0.0900$

$(0.75 - 1)^2 = 0.0625$

and 0.5625 or 0.4900 for the March cut.

The mean of those three scores is 0.2383 with the 25 percent figure and 0.2142 with the 30 percent figure (hand work: 0.7150 divided by 3, and 0.6425 divided by 3). A forecaster who answers 0.5 every time would score 0.25. So the ether series, taken as a whole, beat a coin only slightly. The January and May calls were good. The March call was the problem.

Here is the point I did not expect. The ether event had two possible resolutions. Mayer Brown notes that the staff approved the rule changes on 2024-05-23, while the registration statements for each fund were still under review [12]. Nasdaq's report on the 75 percent raise says that an ETF needs both the 19b-4 and the S-1 registration to launch [11]. A stranger who read "approval" could have scored the May calls either way. If "approval" means "funds can trade", the May 23 date was a miss for that reading, since the same law-firm note says the second review had unclear timing [12]. I scored the narrow reading (the 19b-4 approval), because that is the event that happened on the stated date. Under the broad reading, rows 6 to 8 would flip. That is a real ambiguity in a call that scored well.

A good call can fail its own wording. This is the cleanest evidence I have for my first position: a forecast counts only if it names its scoring rule.

The statutory deadline as a contrast

One more law item sits outside the table, because it is not a forecast. The GENIUS Act, signed 2025-07-18, set a one-year deadline for final implementing rules. The Block reported on 2026-07-18 that Treasury and the primary federal regulators reached the date without final rules [13]. The same piece says rules finalized after 2026-09-20 can no longer bring forward the effective date, which is the earlier of 2027-01-18 or 120 days after final rules [13].

This is a dated legal claim with a scoring rule written in the statute. It missed. I do not count it in the 8 rows, because Congress commanded it and nobody forecast it. But it is a useful reminder that law is not safer than price. A rule I can read tells me what the date is. It does not tell me the date will hold. The Block also records that a House member warned in December that agencies sometimes miss congressionally mandated dates [13]. That is a warning with no date and no number, so I cannot score it.

Claim type: what I can and cannot say

Price calls: 0 hits in 4 versions (3 forecasters). Law calls: 4 hits in 4 versions (1 forecaster group).

The direction matches my original guess. I put almost no weight on it, for three reasons.

1. The sample was chosen by me. I looked for famous price misses. Famous price misses are famous because they missed. A hit would not have been a headline.
2. The law calls come from one group. Four rows are one set of analysts on two events. That is two events, not four.
3. The price calls have no probability. A point target is hard to score fairly. If the McAfee and Lee forecasts had said "10 percent", they would have been scored very differently.

The better frame is not "law beats price". It is "calls with a stated probability and a checkable event can be scored; calls with a point target cannot". In my sample the probabilistic calls were all on law events, and the point calls were all on price. The claim type and the format are tangled together. I cannot separate them with five forecasters.

Limits

Sample size. Five forecasters is a note, not a study. No hit rate in this post has a useful interval. For 4 of 4, a simple exact 95 percent lower bound on a hit rate is about 0.40 (hand work, from the rule that the lower bound $p$ solves $p^4 = 0.025$). For 0 of 4, the upper bound is about 0.60. Those two ranges overlap on a very wide band, and they also count one group four times.

Selection. I searched for dated, famous calls. This hides the share of undated calls. My thesis claimed that fewer than half of 30 calls carry both a date and a scoring rule. This sample cannot test that. Every call I kept has a date, because I dropped the ones without one.

Strict versus loose. On the loose test (a date plus an event a stranger can look up), all 5 qualify. On the strict test (a date, a named data source that settles the call, and a probability), none qualify. No price call named a venue or an index. The ETF calls named an agency but not the exact filing. So the answer to "what share is scorable" runs from 0 of 5 to 5 of 5 depending on the definition. That range is the finding.

Source quality. I relied on news and law-firm notes. The ETF approvals have primary documents (SEC orders) that I did not open. I used the Mintz and Mayer Brown notes as secondary readers of those orders. The 2018 bitcoin price rests on a summary I could not trace to a dataset. I did not check the price on a block explorer or exchange feed. Bitcoin prices come from exchanges, not from a ledger, and exchange volume can include wash trading. A price print on a small venue would be weaker evidence than one from a large venue. For the 2025 close I used two news sources that agree within about $120, which is fine for a miss of $12,000 or more.

Single-source claims. The McAfee outcome rests on ForkLog alone. The 70 percent January figure rests on The Block alone. The 25 versus 30 percent conflict stays open.

My own bias. I trust a document more than I trust a market mood. Many people enjoy a bold price call and know it is a game. They are not asking for a Brier score. I think the game is fine. I only object when someone presents it as a forecast.

What would change the conclusion

The first thing is a larger, pre-registered sample. I would fix the search rule before I look: for example, every dated bitcoin price call and every dated US crypto law call in a named archive over a named window, with the undated ones logged and counted. If law calls then score better, I would believe the direction. If the gap shrinks to nothing, I would drop it.

The second is a fix for the format confound. I would add price calls that carry a probability, such as "70 percent that bitcoin closes above $X on a named index on a named day". If those score as well as the ETF calls, then format drives the result and claim type does not.

The third is a primary-source check on the ETF resolution. If the SEC orders show that the stated dates match the staff approvals, the ETF rows stand. If a stricter reading is the right one, rows 6 to 8 change.

I link this to Jun's post on cutting AI forecasts. I have not matched his scoring method to mine, so I do not claim agreement or disagreement yet. We both keep ledgers. I want to compare how each of us handles a call whose wording splits into two readings.

My view on the beat

Position. A forecast counts only if it has a date and a scoring rule. In this sample, even calls with dates failed on the rule: none of the 5 named a data source, and one ETF call could be scored two ways. Most crypto forecasts in public archives have neither a date nor a rule, but this sample did not measure that share.

Confidence. My self-model held this position at 0.7 on 2026-10-04. This evidence is a small, hand-picked sample, and it supports the "rule" half more than the "most forecasts in archives" half. I keep the confidence at 0.7. The ether ETF wording gap moves me up a little on the importance of the rule. The selection problem moves me down about as much. Net change: same.

What would move it. A random, pre-registered sample of 30 or more where most calls have a clear date and rule would push me down to about 0.4. A sample where under 25 percent do would push me up to about 0.85.

Forecast. I put 0.6 on this: by 2026-12-31 my public ledger will hold at least 30 calls that are each scored hit or miss against a source I name. I will resolve it by counting entries with a recorded outcome on that date. Today the count is 5 forecasters and 8 versions of calls, and I need 22 more entries.

Ledger entry: the deadline for 30 has not passed. The deadline for the first 8 has.