Market moves predictions using Retrieval-Augmented Generation (RAG) analysis of capital market expert opinions in social media

This study explores the predictive value of expert opinions from financial market media using Artificial Intelligence (AI), specifically, Retrieval-Augmented Generation (RAG) framework integrated with a Large Language Model (LLM). By analysing 3,877 YouTube videos spanning 12 months, the research categorised 4,808 expert opinions—either explicit or inferred—into directional market predictions (up, down, flat) for seven diversified financial assets. Results indicate that aggregated expert opinions correlate significantly with short-term market movements but lose predictive power for longer horizons. Explicit opinions demonstrated similar accuracy to inferred judgments, suggesting that LLMs effectively extract latent insights from unstructured data, enhancing accessibility and utility for retail investors. The study highlights the democratising potential of LLMs, providing timely and scalable analysis of vast datasets. However, challenges remain, such as understanding domain-specific nuances and speaker attribution within multimedia content. Statistical analysis reveals that expert opinions, particularly when aggregated, identify exploitable inefficiencies, thereby challenging the Efficient Market Hypothesis's assumption of perfect information dissemination. Short-term market anomalies observed align with behavioural finance theories of cognitive bias and delayed information diffusion. By bridging qualitative sentiment with quantitative modelling, this research underscores the transformative role of AI-driven tools in financial analysis, offering new avenues for individual and institutional investors. Further development of LLMs tailored to domain-specific complexities may revolutionise investment practices and advance research on market behaviour.