78,000-Tweet LLM Study Debunks the Trump-Tweet Market-Mover Myth
A 78,000-tweet LLM study tests whether Trump's posts move markets — and uncovers seven statistical bugs before landing on a null result.
A quantitative researcher tested the trading-desk myth that Trump's social media posts move markets, using Llama-3.3-70B to zero-shot label 78,130 posts and cross-referencing 8,317 market-relevant tweets from a 2025-2026 window against daily OHLCV data for 62 tickers. After pre-registered statistical correction, 0 of 63 test cells showed a significant forward signal — the popular tweet-causes-move story doesn't survive rigorous testing.
The more interesting engineering story is the seven false positives the pipeline produced before reaching that null result: a tie-breaking rule that counted 50/50 splits as wins, a permutation test with collapsed null variance, double-counting bursty tweets against hourly price bars, a hardcoded market-open time that ignored daylight saving, a ticker-matcher that mistook the "President DJT" signature for stock mentions, an unstable return-transform math error, and a keyword filter that conflated "Intel" the company with "intel" the noun — while a private regex quietly deleted counterexamples that undermined the paper's own thesis.
Reversing the question — does the market predict the tweets rather than the other way around — produced a modest, appropriately caveated result: how often Trump posts about a company correlates weakly with the size of a prior price move, though it fails to survive the same multiple-comparison bar applied to the debunked forward signal. The pipeline itself runs on Nebius Serverless: batch AI Jobs and a serving Endpoint share one CPU-only image and call the same decide() function, with the endpoint refusing to boot unless its live prompt hash and data checksum match the batch job's manifest.