Manipulating Headlines in LLM-Driven Algorithmic Trading
The paper demonstrates that Large Language Model-driven trading systems can be manipulated by inserting specific words into news headlines, causing the model to alter its trading decisions. The authors propose a detection mechanism to identify such adversarial modifications.
Background
- This paper shows that AI-powered trading systems using large language models (LLMs) to scan financial headlines can be easily manipulated. Attackers can tweak a headline by swapping a word or adding a negation — changes invisible to a human reader but enough to flip the model's sentiment and trigger a buy or sell order.
- LLMs are increasingly used in finance to automate trading decisions based on news sentiment. This introduces a new vulnerability: adversarial text attacks, where crafted inputs fool the model without looking suspicious to people.
- The research bridges two areas: adversarial ML (finding inputs that trick AI models) and market manipulation (e.g., "pump and dump" schemes). It warns that LLM-driven trading faces unique security risks that purely numerical trading algorithms did not.