Researchers have combined large language models with complex geometry to create a new “emotional vulnerability” metric that outperforms traditional sentiment analysis.
A new study from the Chinese Academy of Sciences suggests that predicting stock market crashes isn’t just about knowing if investors are happy or sad – it’s about measuring how “fragile” those emotions are.
In a paper published in Risk Sciences, researchers from the Central University of Finance and Economics in Beijing unveiled a novel method that combines artificial intelligence with network science to identify hidden risks in the stock market.
The team proposes a new concept called emotional vulnerability. Unlike standard sentiment analysis, which only tracks whether the market is optimistic or pessimistic, this new metric captures the stability and structural integrity of those feelings.
Geometry of the market
To calculate this, the researchers analysed millions of stock forum posts from China’s A-share market.
They used an advanced large language model (LLM) designed for Chinese text – the Moka Massive Mixed Embedding (M3E) model – to convert vast amounts of investor chatter into numerical vectors.
These vectors were then used to build “sentiment networks” for individual stocks. By applying a geometric concept known as Ricci curvature, the team could measure the “tightness” of these networks.
“The vulnerability of each sentiment network was measured using Ricci curvature, a concept from geometry that has been widely applied in financial system risk analysis,” explained corresponding author Ning Zhang, a professor at the Central University of Finance and Economics.
“In this context, higher curvature indicates tighter connections and more homogeneous investor sentiment, making markets more fragile when disrupted”.
Profiting from stability
The study’s results were stark. When emotional vulnerability was high – meaning investor sentiment was tightly packed and uniform – stock prices were prone to falling or becoming unstable.
Conversely, when vulnerability was low, markets showed stronger resilience and generated higher excess returns.
Crucially for investors, the team found that portfolios composed of “emotionally stable” stocks consistently outperformed those composed of stocks with high emotional vulnerability.
“This relationship remains significant even after controlling for traditional asset pricing factors… as well as for sentiment direction (optimistic or pessimistic),” Zhang added.
The authors suggest that by shifting attention from the direction of sentiment to its structure, this new tool could significantly improve risk assessment and help investors make more informed decisions.
