Tracing Knowledge Flows in Financial Markets with Hybrid Sentiment AI

Abstract

Financial markets are commonly described as information-efficient, yet the phrase conceals more complexity than it clarifies. Between the arrival of information and the adjustment of prices lies a process that is neither fully observable nor mechanically uniform. Interpretation mediates how signals are received, credibility shapes their perceived relevance, and responses unfold unevenly across participants. From a knowledge management perspective, this raises a central question: how does dispersed information consolidate into knowledge that meaningfully informs decisions? The framework we developed does not seek to model the entire market mechanism. We concentrate instead on a measurable segment of that transformation. Media sentiment is examined using a hybrid analytical approach that combines the lexicon-based VADER model with the domain-specific transformer model FinBERT. The two models do not always converge, and this divergence provides additional insight rather than methodological weakness. Their integration enables us to capture both surface polarity and contextual financial nuance within a single structure. Because informational inputs differ in reliability, sources are weighted by credibility and filtered through an entity-based relevance mechanism. Sentiment scores are aggregated across adjustable time windows, enabling us to detect whether informational pressure accumulates before a price adjustment. Using selected S&P 500 firms characterised by sustained media exposure and high liquidity as the empirical setting, we conduct structured backtesting to evaluate directional alignment and interval coherence across multiple horizons. The main objective of this paper is to assess whether consistent relational patterns can be identified. Our results indicate that such patterns are present, although their strength varies across horizons and informational density. In some instances, aggregated sentiment precedes adjustment, while in others it dissipates without measurable impact. This unevenness suggests that sentiment analysis may be more appropriately understood as a mechanism for tracing knowledge flows within financial ecosystems rather than as a deterministic forecasting device. From this perspective, our study contributes to knowledge management by offering a computational approach for examining how dispersed information becomes actionable knowledge in high-velocity environments.