Sentiment Analysis and Large Language Models in Foreign Exchange Forecasting: A Bibliometric Perspective

Foreign exchange forecasting remains a demanding research area because prices respond quickly to news, policy communication, and shifts in market expectations. The scientific literature has shifted from macroeconomic and econometric models to machine learning and deep learning, with increasing attention to textual information and sentiment as complementary inputs. Recent studies confirm that sentiment features improve short-horizon forecasting, though traditional methods struggle with context, negation, and domain-specific language.

This paper applies a bibliometric analysis to map how research at the intersection of foreign exchange prediction, sentiment analysis, and advanced language modeling has evolved. Bibliographic records were retrieved from the Web of Science database, based on 48 identified publications, with 34 studies published from 2019 onward selected for analysis. The study uses VOSviewer to identify co-authorship networks, cocitation structures, keyword co-occurrence clusters, and emerging thematic trends. The research asks how the literature structures the role of sentiment in exchange rate forecasting, how context-aware language representations are positioned relative to traditional sentiment pipelines, and which integration patterns are most common when combining textual signals with numerical time series.

The results show distinct research clusters linking forecasting architectures, financial text processing, and sentiment-based modeling, as well as a growing focus on context-aware language representations. The analysis also highlights persistent gaps related to end-to-end integration of text semantics, timing alignment between text and market data, and consistent evaluation framing. The paper maps the field and proposes research directions for more coherent sentiment-augmented forecasting designs.