As advancements in artificial intelligence continue to accelerate, new opportunities are emerging for integrating cutting-edge techniques into the investment process. Large language models (LLMs) have shown remarkable promise in parsing qualitative data, uncovering insights, and generating signals that can inform stock selection.
At Context Analytics, we specialize in transforming unstructured data from alternative sources into actionable market intelligence. One of the richest and most dynamic of these sources is X (formerly Twitter), where market participants share real-time opinions, reactions, and news. By capturing and structuring this conversation, we provide investors with unique perspectives that can complement traditional quantitative models.
Our pipeline begins with comprehensive coverage of all relevant tweets for each security. To convert this flood of information into a more digestible format, we developed the AI Summary Feed. Every 15 minutes, our system condenses the prior 24 hours of Twitter conversation around each security into a concise, qualitative summary.
This feed is designed to provide investors with fast, contextual insights into what the market is saying. But it also presents an opportunity: what if we could transform these summaries into trading signals?
To test this, we focused on the AI Summary closest to market close each day and created a rolling 3-day (72-hour) concatenated summary for every security. This gave us a consolidated view of recent market conversation leading into the next trading day.
From there, we introduced an additional LLM decision layer. Using the following structured prompt:
The LLM was asked to analyze the 72 hours of summaries and issue a simple 24-hour recommendation: Buy, Neutral, or Sell.
We back tested this strategy starting from January 1, 2025, across the universe of U.S. securities with more than five tweets in the prior 24 hours. Each day, we constructed portfolios for Buy, Neutral, and Sell signals, trading them daily close to close and comparing results against the SPY benchmark.
The findings were striking:
This asymmetry reveals something important: LLMs, when analyzing summaries generated by other LLMs, tend to lean positive or neutral. But when the model is confident enough to issue a Sell signal, it captures a truly negative consensus that translates into a powerful short opportunity. Our back test results provide evidence
Several factors make this research novel:
This experiment underscores the potential of combining high-quality alternative data from Context Analytics with advanced LLM reasoning to uncover alpha opportunities. While early results show that long and neutral signals may lag the market, the outsized performance of the short signals offers a compelling case for further research and refinement.
At Context Analytics, we remain committed to exploring these frontiers—pushing the boundaries of how AI, unstructured data, and market intelligence intersect.
Stay tuned for more updates on this research, and for more information, visit contextanalytics-ai.com.
TL;DR:
Using AI-generated summaries of Twitter conversations, Context Analytics built a strategy where LLMs issue daily Buy, Sell, or Neutral signals. While Buy and Neutral baskets underperformed, the Sell signals produced outsized short alpha: a 40% YTD drop in selected names vs. SPY up 10%, translating to a 30% long/short spread. This shows LLMs can surface rare but high-conviction downside opportunities from qualitative social data.