AI-generated research, checked but not peer-reviewed. It may contain errors.

E904　 KILL　 UnderpoweredTested · October 2026

After a head of state's posts, crypto shows no reliable follow-through

Entering crypto two minutes after a head of state's keyword post gained +67.9 bp (BTC) in 2025 at t 1.68 against a gate of 2.94; the same rule returned -5.8 bp in 2026 and no instrument passed.

event-study news political-posts bitcoin ethereum gold-token continuation unscheduled-news

Readable summary Full research record JSON recordFeedback on E904

Posts by a head of state on their own social platform move markets, and crypto trades around the clock, so some believe the move keeps going for hours after a slow reader reacts. We took 903 keyword posts (tariffs, China, the Fed, bitcoin and others) from 2025-2026 and entered two minutes after each post that moved the price by at least 30 bp (15 bp for the gold token). In 2025, Bitcoin continued by +67.9 bp over 4 hours (29 events), but the t-statistic was 1.68 against a gate of 2.94. In 2026 the same rule returned -5.8 bp (42 events). No instrument passed the selection gate.

Bitcoin, 2025 selection+67.9 bp4 h continuation, n 29, t 1.68; gate: t 2.94 and n 40

Bitcoin, 2026 held-out-5.8 bpn 42, t -0.38; the held-out step was not formally run

Ethereum, 2025 selection+41.9 bpn 56, t 0.67, worst event -1,177 bp

Volatility effect67 vs 60 bpmean 4 h absolute Bitcoin move after post clusters vs unconditional

01 · The idea being tested

When a head of state posts about tariffs, China, the Fed or crypto, markets react slowly, so a trader who reads the post a couple of minutes late can still ride the move.

Why people believe it

People argue that policy news arrives as a post at an unscheduled time, often outside US stock-market hours, so crypto is the only liquid market open and absorbs it first. Text needs interpretation, so slower participants and trend-following or risk-parity funds react over hours rather than seconds.

Why it would persist The argument, as stated before testing

The posts are unscheduled, need judgement to read, and many are noise, so a mechanical rule is hard to build and the signal is hard to arbitrage away. The idea fails if the whole reaction is finished within two minutes because bots read the feed.

02 · How it was tested

We took a public archive of one head of state's posts on their own social platform from late January 2025 to 2026-10. We kept original posts that contained a fixed list of 17 keywords and merged posts within 60 minutes of each other. For each post we measured how far Bitcoin, Ethereum and a gold-backed token moved from one minute before the post to two minutes after it. When that move was large (30 bp or more for the first two, 15 bp for the gold token), we pretended to enter at the two-minute mark in the direction of the move and recorded the result 15 minutes, 1 hour and 4 hours later. 2025 chose the rule; 2026 was reserved to confirm it.

Pass required

Selection, per instrument: at least 40 reaction events, absolute mean C4 of at least 35 bp, median of the same sign, day-clustered t of at least 2.94 (Bonferroni for 3 instruments, two-sided 5%) and event-minus-control of the same sign. Held-out: same sign, absolute mean of at least 25 bp, at least 25 events, day-clustered t of at least 2.0.

Killed if

No instrument passes selection, or the held-out step fails.

Costs

A taker round trip on a perpetual at a retail crypto venue was assumed at about 12 bp, which set the required gross mean at 35 bp in absolute value. The results below are gross.

Controls

Event minus control. For each reaction event, 5 control minutes were drawn at random from non-post minutes in the same hour of the week within 8 weeks, with the same absolute-move threshold and no qualifying post within 6 hours; the control outcome is the 4 h continuation of those generic one- to three-minute jumps. Also reported: the 4 h absolute move after any post cluster against the unconditional 4 h absolute move.

03 · What happened

Why it was rejected

No instrument passed the selection gate: Bitcoin had 29 events (40 required) and t 1.68, Ethereum t 0.67, and the gold token 13 events and a negative mean (-20.5 bp), against a required t of 2.94.

The 903 clusters of original keyword posts yield 240 reaction events in total: 71 for Bitcoin, 113 for Ethereum and 56 for the gold token, counting both periods.
Bitcoin selection (2025): 29 events, 4 h continuation +67.9 bp gross (median +47.9 bp, t 1.68). The control jumps continued by +23.8 bp, leaving +44.1 bp over control. The mean is above the 35 bp line, but n and t are below the gates, so the gain is not distinguishable from noise.
Ethereum selection (2025): 56 events, +41.9 bp gross (median +48.3 bp, t 0.67). The worst single event lost 1,177 bp, so the average depends on a wide spread of outcomes.
Gold token selection (2025): 13 events, -20.5 bp. Its sign is opposite to the other two, and 13 events are too few to say more.
Held-out 2026, reported for information only because no instrument passed selection: Bitcoin -5.8 bp (42 events, t -0.38), Ethereum +3.9 bp (57 events, median -6.5 bp). The 2025 continuation in Bitcoin and Ethereum is absent in 2026.
Post clusters raise the 4 h absolute move only slightly: 67 bp against 60 bp unconditionally for Bitcoin, 102 against 92 bp for Ethereum and 37 against 31 bp for the gold token.
Bitcoin and Ethereum continuation grew with the horizon in 2025 (Bitcoin +24.0 bp at 15 minutes, +44.3 bp at 1 hour, +67.9 bp at 4 hours) but was within -7.0 to +3.9 bp at every horizon in 2026. The growth did not hold and was not distinguishable from noise. The 95% bootstrap intervals of daily means for Bitcoin 2025 are -5.7 to +200.1 bp and for Bitcoin 2026 -55.4 to +39.0 bp.

4-hour continuation: every interval includes zero

Mean direction-adjusted 4 h return after a +2 min entry, gross, with 95% bootstrap intervals of daily means. Every interval spans zero. Bitcoin and Ethereum are positive in 2025 and about zero in 2026; the 35 bp line is a gross threshold set from an assumed 12 bp round-trip cost, not a net profit.

Chart data (bp)

	Mean	95% low	95% high
BTC 2025 (n 29)	67.9	-5.7	200.1
BTC 2026 (n 42)	-5.8	-55.4	39
ETH 2025 (n 56)	41.9	-60.8	130.5
ETH 2026 (n 57)	3.9	-47.7	43.7
Gold token 2025 (n 13)	-20.5	-50	6

The t-statistics never approached the gate

2025 selection2026 held-out

Day-clustered t of the 4 h continuation by instrument and year. The selection gate was 2.94; the best 2025 value was 1.68 (Bitcoin). In 2026 the Bitcoin and Ethereum values are near zero (held-out figures are information only).

Chart data (day-clustered t)

	2025 selection	2026 held-out
Bitcoin	1.68	-0.38
Ethereum	0.67	-0.05

The 2025 follow-through is gone in 2026 at every horizon

BTC 2025BTC 2026ETH 2025ETH 2026

Mean direction-adjusted return after a +2 min entry at 15 min, 1 h and 4 h, gross. Bitcoin and Ethereum rise with the horizon in 2025 (up to +67.9 and +41.9 bp at 4 h); in 2026 all six values lie between -7.0 and +3.9 bp.

Chart data (bp)

	BTC 2025	BTC 2026	ETH 2025	ETH 2026
15 min	24	-6.1	18.6	-0.6
1 hour	44.3	-2	43.2	-7
4 hours	67.9	-5.8	41.9	3.9

04 · What it means

Why it failed Interpretation · not established

The 2025 means for Bitcoin and Ethereum were never distinguishable from noise: with 29 and 56 events and single outcomes in the hundreds of bp, Bitcoin's t of 1.68 stayed below the 2.94 gate and Ethereum's standard error was larger than its mean. The 2025 to 2026 change therefore cannot be called a decay of an established effect; the 2025 estimate may simply have been one favourable draw. Whether the 2025 sample was dominated by one policy theme, and whether markets absorb these posts faster now, are explanations the data do not test.

Lesson

A mean that clears the cost line is not evidence when the sample is a few dozen events with a standard deviation of hundreds of bp. For rare unscheduled news the sample is small by construction, so the test needs a significance gate fixed in advance and a later period that was held back; here the gate stopped the rule, and the held-out period was not formally run because selection failed.

Scope of this result

Covers one account's keyword posts, late January 2025 to 2026-10, three spot pairs, a notional entry two minutes after the post and exits at 15 minutes, 1 hour and 4 hours, gross of costs. The keyword list and the 30 bp and 15 bp thresholds were fixed in advance. It does not cover other accounts, other keywords or other entry delays.

A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.

1. Test design

Pre-registered

Yes: rule and criteria fixed before the test data was examined

Data

A public archive of one head of state's posts on their own social platform: post id and timestamp; only keyword flags were kept — a news organisation's public data archive; per post, millisecond timestamps; 2022-02..2026-10 (test uses 2025-01..2026-10); single public download, no login | 1-minute closing prices of BTC/USDT, ETH/USDT and PAXG/USDT spot pairs — the public monthly and daily archives of the largest crypto spot venue; 1 minute (closes only); 2025-01..2026-10; public download, no login

Universe

Bitcoin, Ethereum and a gold-backed token (PAXG), each against the dollar-pegged USDT, spot.

Signal

Original posts (reposts dropped) from late January 2025 to 2026-10 whose lower-cased text contains one of: tariff, china, chinese, a foreign leader's name, iran, sanction, fed (with trailing space), federal reserve, the central bank chair's name, interest rate, oil, bitcoin, crypto, strategic reserve, trade deal, trade war, retaliat. A post within 60 minutes of an earlier qualifying post is merged into it (the first counts). The whole archive (2022 onward) holds 37,347 posts, including 5,591 reposts and 1,870 keyword posts; the original keyword posts from late January 2025 merge into 903 clusters; 21% fall on weekends. Reaction M = P(t+2 min) / P(t-1 min) - 1, with t the minute of the post. A reaction event needs |M| of at least 30 bp (BTC, ETH) or 15 bp (PAXG). Outcome C4 = sign(M) x (P(t+4 h) / P(t+2 min) - 1), with C15 and C60 measured to +15 minutes and +1 hour.

Execution

Notional entry at the 1-minute close two minutes after the post, in the direction of the first move; exit 15 minutes, 1 hour or 4 hours later. All results are gross, before costs.

Costs

A taker round trip on a perpetual at a retail crypto venue was assumed at about 12 bp, which set the required gross mean at 35 bp in absolute value. The results below are gross.

Controls

Sample split

Selection 2025-01..2025-12, held-out 2026-01..2026-10. The held-out step runs only for an instrument that passes selection.

Pass criterion

Kill criterion

No instrument passes selection, or the held-out step fails.

Power

Not computed in advance. From the reported selection figures, the day-clustered standard error was about 40 bp for Bitcoin (67.9 / 1.68) and 63 bp for Ethereum (41.9 / 0.67), so a mean would have needed to be roughly 119 bp and 184 bp to reach t 2.94. The 35 bp cost line was far below what 29 and 56 volatile events could resolve.

2. Result

Observations

240

Instruments

Net mean

—

95% interval

Not reported

Median

—

Worst

—

Finding 1

The 903 clusters of original keyword posts yield 240 reaction events in total: 71 for Bitcoin, 113 for Ethereum and 56 for the gold token, counting both periods.

Finding 2

Bitcoin selection (2025): 29 events, 4 h continuation +67.9 bp gross (median +47.9 bp, t 1.68). The control jumps continued by +23.8 bp, leaving +44.1 bp over control. The mean is above the 35 bp line, but n and t are below the gates, so the gain is not distinguishable from noise.

Finding 3

Ethereum selection (2025): 56 events, +41.9 bp gross (median +48.3 bp, t 0.67). The worst single event lost 1,177 bp, so the average depends on a wide spread of outcomes.

Finding 4

Gold token selection (2025): 13 events, -20.5 bp. Its sign is opposite to the other two, and 13 events are too few to say more.

Finding 5

Held-out 2026, reported for information only because no instrument passed selection: Bitcoin -5.8 bp (42 events, t -0.38), Ethereum +3.9 bp (57 events, median -6.5 bp). The 2025 continuation in Bitcoin and Ethereum is absent in 2026.

Finding 6

Post clusters raise the 4 h absolute move only slightly: 67 bp against 60 bp unconditionally for Bitcoin, 102 against 92 bp for Ethereum and 37 against 31 bp for the gold token.

Finding 7

Bitcoin and Ethereum continuation grew with the horizon in 2025 (Bitcoin +24.0 bp at 15 minutes, +44.3 bp at 1 hour, +67.9 bp at 4 hours) but was within -7.0 to +3.9 bp at every horizon in 2026. The growth did not hold and was not distinguishable from noise. The 95% bootstrap intervals of daily means for Bitcoin 2025 are -5.7 to +200.1 bp and for Bitcoin 2026 -55.4 to +39.0 bp.

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Download the public archive of the account's posts (id and timestamp are enough) and drop reposts. 2) Flag posts whose lower-cased text contains the 17 keywords, keep late January 2025 to 2026-10, and merge posts within 60 minutes of an earlier kept post. 3) Download the 1-minute spot closes for BTCUSDT, ETHUSDT and PAXGUSDT from the venue's public archives and forward-fill gaps. 4) Per post, M = P(t+2m) / P(t-1m) - 1; keep |M| of at least 30 bp (15 bp for PAXG). 5) C4 = sign(M) x (P(t+4h) / P(t+2m) - 1), also to +15 min and +1 h. 6) Draw 5 control jump minutes per event (same hour of the week, within 8 weeks, no qualifying post within 6 hours). 7) Compute the mean, median and day-clustered t per instrument for selection and held-out, and check the gates.

Grade

COMPLETE