Everyone Knows Trump's Tweets Move Markets. I Measured It: the Connection Is Real — the Direction Isn't.

#NebiusServerlessChallenge

Live dashboard: Myth-Busting Quantitative Terminal
· Live endpoint: /predict on Nebius Serverless
· Code: github.com/KoralZakai/stocksPredictionAfterTweet

Everyone on a trading desk knows the story. He tweeted about Intel, and the stock ran for
months. He posts about Iran, oil spikes. The anecdotes are vivid, specific, and everybody
has one.

While the prevailing myth suggests that Trump's tweets drive market movements, my research —
built on an analysis of 78,130 posts, with Llama-3.3-70B reading every market-relevant tweet —
reveals a more nuanced reality. The findings demonstrate that the market tends to price in
significant events long before the tweet is even posted. This indicates a case of reverse
causality, where the market dictates Trump's narrative, rather than the other way around. He
isn't moving the market; he is simply riding the wave of established trends.

Forward — the direction you could trade — the signal is a coin flip, and I can show that with
pre-registered tests rather than vibes: 0 of 63 cells survive correction. Getting to that
answer honestly was the hard part. Seven times this pipeline produced a beautiful,
publishable, completely false positive — each one looking exactly like the discovery the
anecdotes promise. Those seven are the engineering content of this post.

---

The setup

Data. 78,130 public posts; 8,317 in the study window (2025-01-01 → 2026-07-06). Daily
OHLCV bars for 62 tickers, committed to the repo. Public data only — a stranger can re-run
every number without a single private key.

The model. meta-llama/Llama-3.3-70B-Instruct on Nebius AI Studio, zero-shot,
reading only the tweet text: what is this post about, which instruments does it touch, and
which way does each one go. 476 tweets, 1,296 instrument calls. No fine-tuning — the question
is whether the text carries signal, and a fine-tune would smuggle the outcome into the
answer.

The architecture. Nebius Serverless AI Jobs run the batch research pipeline; a Serverless
AI Endpoint serves /predict. One CPU-only image builds both, so batch and serve cannot
drift. The rule that makes this rigorous rather than plumbing: the serverless layer
contains zero science. Every job is a thin CLI that calls pure modules and writes
artifacts. The endpoint is a thin handler that calls the same decide().

Jobs:      data_ingestion → labeling → features → dataset → training → evaluation → reporting
Endpoint:  POST /predict ──── the same decide() ────┘
Feedback:  /predict → prediction_log.jsonl → jobs/feedback.py → replication report

The endpoint refuses to boot if its live classification prompt hash or its corpus
sha256 don't match the manifest the batch job produced. The Job produces the numbers; the
Endpoint may only cite them.

---

The seven artifacts

1. The tie-break that counted ties as wins

_tweet_hit aggregates a tweet's instrument basket into one verdict. The tie rule:

return int(sum(hs) / len(hs) >= 0.5)   # 1-of-2 hits → 0.5 >= 0.5 → a WIN

49% of baskets hold exactly two instruments, where "majority" is undefined. Every 50/50
split scored as a full win. Headline hit rate: 0.618, p=0.017. I had a finding.

A tie carries no directional verdict — it's unscoreable, not a win. Excluding ties: 0.507.
Scoring them as losses is the mirror bug (0.393). Two independent unbiased estimators agree
at ~50%. The entire "signal" was the tie rule.

2. The permutation null that couldn't say no

The event study flagged 22 of 72 cells as significant. Gorgeous. Also arithmetic: he
posts on 83% of all trading sessions, so sampling a null pool of size ≈ the population
without replacement collapsed the null's variance and floored every p-value. Bootstrap with
replacement + dedupe to unique event-days: 22 → 0.

3. The same-bar burst

The intraday study returned 9 solid cells, p≈0.0018 — "war tweets spike oil volume 79%".
He posts in bursts: five tweets in ten minutes. With one price bar per hour, all five pointed
at the same move. One event counted five times as five pieces of evidence.

I only caught it because the "biggest movers" list had a tweet about a renovated Palm Room
at the top. Count each hour once: 9 → 0.

4. A hardcoded market clock

13.5 for the US open. NYSE opens 13:30 UTC in EDT but 14:30 UTC in EST — so every tweet
posted 13:30–14:30 UTC from November to mid-March was anchored to the wrong session.

5. DJT — his own signature

Trump signs posts "President DJT". The ticker rule read that as a mention of his listed
company: 754 of 762 "DJT mentions" were the sign-off.

6. Negated-return math

-E(-X) to get a backward return. Valid for small moves; on a stock that had doubled it
reported +44.8% for a +105.3% run-up.

7. "intel" is not Intel

The one that stings. In a political corpus "intel" overwhelmingly means intelligence:
51 of 55 matches (93%) were spy-agency tweets — leaking intel, Danish intel, Obama
Intel Chief — and they were sitting in the CORPORATE cohort of a registered study.

But the mirror half is the lesson. My dashboard kept its own private Intel regex that
only accepted company context after the word. So it scored these as not Intel:

• The CEO of INTEL is highly CONFLICTED and must resign
• I met with Mr. Lip-Bu Tan, of Intel
• I PAID ZERO FOR INTEL, IT IS WORTH APPROXIMATELY 11 BILLION DOLLARS

Those are exactly the mentions that contradict my own headline exhibit. Intel was
falling (−14.3%) when he demanded the CEO resign — then rose +18.1%. My exhibit claimed
"he reflects news rather than generating it," and my regex had quietly deleted four
counterexamples out of eight.

One rule dropped 93% true positives. The other dropped the inconvenient cases. A rule that
keeps what fits the thesis is the anecdote generator I was writing about — running on my own
exhibit. Fixed, rescored: CORPORATE 43 → 13 events, 72 → 63 cells, verdict unchanged: 0
survive.

---

Which way does the arrow point?

The last thing I built asks the connection question directly, on the same committed data.
Not "do his tweets predict prices?" — that's dead — but "does the market predict his
tweets?" Three measurements, in the order they survived (experiments/event_study/mirror_test.py,
seeded, offline):

1. Tweet level. How big was the abnormal move a mentioned company had behind it, vs random days? 9.7% vs 6.7%, p<0.001. A finding! — and it's fake. It's the burst artifact again: ten of those "observations" are ten tweets about the same Tesla fall. My own artifact #3, back for an eighth attempt.
2. Episode level (one observation per company × 21-session window): the gap shrinks to 8.1% vs 7.1%, p=0.17. Dead. He does not systematically pick movers.
3. Intensity. But how many times he posts about a company tracks the size of the move behind it: episodes he mentions once sit on ~4% moves; episodes he posts about three or more times sit on ~8.4% (rank correlation +0.23, raw p=0.033, n=65). He doesn't tweet about more companies when markets move — he tweets more times about the same one.

Honest label: that last result is exploratory — one test, raw p, and if I apply the
same multiple-comparison correction to the mirror script's own three tests that killed the
forward signal, p≈0.10 and it fails too. I'm reporting it as suggestive, not as a survivor;
holding my favourite direction to the same bar as the one I rejected is the whole point of
this project. What doesn't need a p-value is the case-level record above: the market
moved first, in every famous anecdote I could check.

The timing is the whole illusion. His posts land when a stock is already deep inside
a strong, established trend — Intel had gained +105% over the prior 21 sessions when
he posted "Intel Stock continues to rise"; Tesla was mid-fall when the thirteen Tesla
posts arrived. To an observer who sees tweet-then-more-movement, he looks like the cause.
But the trend was there first: he is reflecting an established market move, narrating it
— and memory quietly reorders the sequence into tweet-then-move. That is why the anecdotes
feel irrefutable and why every one of them dissolved under a timeline: the observed
correlation is real, the implied causation is backwards.

So the honest summary of the connection: his feed is a mirror of the market, not a
lever on it — and the mirror hypothesis deserves its own pre-registered study, which
the repo is set up to run.

The results

➡ Live dashboard: Myth-Busting Quantitative Terminal
— every number below is explorable there, including the wrong-day control and the
per-tweet reasoning behind each call.

test	result
EOD directional (tweet-level)	0.522, n=67, p=0.40 → shipped_horizons: []
Beat-SPY (instrument-level, 1,296 calls)	48.1% / 47.3% / 48.6% / 49.2% at EOD/3d/1w/1mo
Event study, 63 pre-registered cells	0 survive BH (min p_bh 0.76)
Intraday shock, 30 cells, MDE 0.1–0.3%	0 survive
Meta-model ("when can we trust it?")	Val AUC 0.593 → Test 0.431

Two tells explain the coin flip. 73% of the model's calls are "UP" — it isn't reaching a
direction, it mostly says up, and markets drift up anyway. And its most-named instrument is
SPY itself (304 calls), which cannot beat itself.

Then the control that settles it. The model's named assets do move more than the ones it
ignores (1.52% vs 1.12% at one day). So I scored each tweet's picks on a different tweet's
day: +0.40pp real vs +0.38pp wrong-day. The entire advantage survives on a day the
tweet had nothing to do with. It isn't knowledge — it's that oil and the fear gauge are
jumpier than staples and bonds every day of the year.

Why the anecdotes survive contact with all of this: reverse causality (the mirror,
above), mean reversion (big oil moves recover the same with no tweet), and saturation (he
posts on 83% of sessions, so every big move has a coincident tweet available to blame).

Closing the loop, honestly

/predict appends every served call to a bucket-mounted log; a Job waits for each horizon to
close, labels it, and scores it against the same pre-registered registry.

The obvious version of this — retrain on live calls, ship when p < 0.05 — is the fastest way
to destroy everything above. Callers choose which tweets to send, so that sample is
selected, not sampled; and re-scoring a growing log until it passes reaches significance
with probability 1 on pure noise. So live calls are a replication set, never training
data, and the correction is applied over registry × n_looks — the bar rises every run.
You cannot wait your way to a finding.

What shipped

An endpoint that abstains on every tweet — because nothing survived correction, and a
calibrated refusal is the correct output. It quotes no accuracy and no confidence score,
because nothing I measured earned one. Try it:
/predict
— send any tweet and it will tell you, with its reasoning, why it will not call it.

Serverless fits research better than I expected: every stage is deterministic, containerized,
hash-stamped, independently re-runnable, CPU-only, and scales to zero. But the real lesson is
cheaper than infrastructure. Six of my seven artifacts were caught by a control, not by
inspection — and the seventh was caught because a tweet about a renovated Palm Room
topped the leaderboard.

Build the thing that can say no. Then check whether it's saying no for the right reason.