Using social media sentiment analysis to predict stocks

I was speaking to a physics PhD student from the UK who proposed that more institutional traders could potentially gain an edge over retail traders by using natural language processing to extract sentiment from social media. They would then train machine learning models on historical data to identify when sentiment shifts toward particular companies correlate with subsequent price movements.

While catastrophic examples are sobering, they illustrate the concept: after the UnitedHealthcare CEO shooting in December 2024, UHC's stock price continued to move over the following trading sessions as the story developed. Similarly, Silicon Valley Bank's collapse in March 2023 was accelerated (though not solely caused) by sentiment spreading rapidly on social media, which triggered a classic bank run as depositors rushed to withdraw funds.

The approach isn't traditional quantitative finance based on financial statements or technical indicators. Instead, it's sentiment analysis: using NLP models (including LLMs) to process social media text, quantify sentiment, and potentially identify early warning signs of price movements. The idea is that large scale sentiment signals might show up before prices fully adjust, particularly when there's a gap between when news breaks and when all market participants react to it.

Key limitations: correlation doesn't mean reliable prediction, markets are increasingly quick at pricing in sentiment, and there are significant regulatory questions around using certain types of data.