Sentiment Correlation in Financial News Networks and Associated Market Movements

• Complex interdependencies in financial markets motivate data-driven network approaches, as classical economic models have often been inadequate for explaining lower-level market dynamics.Source: Introduction · p. 1
• The work addresses a gap between network-based market dynamics and NLP-based financial-news sentiment, previously studied largely as separate directions.Source: Introduction · p. 1
• The study combines NLP sentiment analysis with a news-based network of 87 companies, drawing Reuters articles from 2007 to 2013.Source: Introduction · pp. 1–2
• Community detection on the 87-company news network yields groups often closely aligned with conventional sector classifications.Source: Introduction · p. 2
• Sentiment events are often correlated with sentiment movements among related-company groups and are associated with unusual return and volatility movements at company and group levels.Source: Introduction · p. 2

Evolution of Sentiment

• Company sentiment varied across the study: U.S. banks showed strong negative sentiment around the 2008 crisis, while Technology sentiment was relatively stable and Financial Services more volatile.Energy and Financial Services sentiment was negative in most quarters.Source: Evolution of Sentiment · p. 2

News Co-occurrence Network

• 0.0157 versus 0.00229: median co-occurrence weight is higher for in-sector than out-sector company pairs, and detected groups broadly match conventional sectors.The network uses company co-mentions to define weighted links and community detection to identify groups.Source: News Co-occurrence Network · p. 2
• The network also reveals cross-sector business relationships, while quarterly centrality and network structure change over time.Detected groups can unite companies placed in different Bloomberg sectors; network average degree peaked in late 2008, and group similarity to sectors generally increased over the study period.Source: News Co-occurrence Network · p. 4

Sentiment Dynamics in Related Companies

• Sentiment events in one company correlate with changes in related-company group sentiment, with negative events often producing larger movements and polarity symmetry varying by group.Financial-dominated Group 1 shows larger group movement after negative events, while consumer-related Groups 2 and 4 show more symmetric responses to positive and negative events.Source: Sentiment Dynamics in Related Companies · pp. 4–5, Figure 2 · p. 5

Market Movements Associated with Sentiment

• Across most groups, sentiment events are associated with post-event CAR divergence and elevated volatility; positive events align with rising prices and negative events with declines.For some groups, the market response continues for several days after the event.Source: Market Movements Associated with… · pp. 5–6, Figure 3 · p. 6, Figure 4 · p. 7
• At the group level, extraordinary sentiment changes coincide with shifts in abnormal-return distributions and elevated volatility, including in groups with weak individual-company associations.For Industrial-dominated Group 2, abnormal-return probability mass shifts opposite to the direction of group sentiment change.Source: Market Movements Associated with… · p. 6

Discussion

• The proposed network pipeline links financial-news sentiment with company inter-connectedness and identifies associated market movement beyond the individual-company level.The authors describe this as a first step and a practical use of the network representation.Source: Discussion · p. 7
• Considering company or asset interplay and shifting analysis toward sector or group level may support more informed investment decisions and systemic-risk monitoring.The authors present these as potential benefits for market participants and regulators.Source: Discussion · p. 7
• The study is limited to Reuters coverage and the largest companies with adequate Reuters coverage, and its NLP depends on entity recognition and sentiment prediction that remain open research questions.Source: Discussion · p. 7
• Figure 5 presents probability-density profiles of abnormal return and daily realised volatility for Group 2 around group-aggregated negative and positive sentiment events.Source: Figure 5 · p. 8
• The analysis uses historical data and does not test causality; it also leaves overlapping sentiment events from different companies uncontrolled.Source: Discussion · p. 8

Data sets

• Reuters coverage from 2007–2013 and Bloomberg daily market data provide the news and price series for 87 target companies.The news sample covers 27 full quarters, from January 1, 2007 to September 30, 2013.Source: Data sets · p. 8

Named Entity Recognition

• Named-entity recognition retains organisation mentions, which are normalized and narrowed to 87 Bloomberg-tickered firms; Louvain communities and an IQR rule identify related-company clusters and outlier pairs.The IQR rule flags edge weights above the 75th percentile plus 1.5 IQR in both within-sector and out-of-sector categories.Source: Named Entity Recognition · p. 9

Sentiment Prediction

• A context-aware sentiment model assigns each company mention a score from -1 to 1; mentions are averaged into article-level and company-level time series.Group sentiment is the average of constituent companies’ scores.Source: Sentiment Prediction · pp. 9–10

Sources of Sentiment

• Sentiment is concentrated in a small article subset: the top 9.2% account for half, and manual review found little market commentary in sampled high-sentiment articles.The top 20 articles were inspected, with the pattern also observed in the top 50; the authors also examined the five strongest-sentiment days for three representative companies.Source: Sources of Sentiment · p. 10

Sentiment Event Days and Group Aggregate Sentiment

• Events exceed two standard deviations above or below the preceding 180-day sentiment average; five factors explain 2.3% of event-series variance versus 55% of daily sentiment.Event-aligned sentiment is compared with dynamically selected top-10 network neighbours over a 60-trading-day news window, then averaged across event days and company groups. A typical-sample review suggests event-day sentiment mainly reflects fundamental events rather than market commentary.Source: Sentiment Event Days and Group… · pp. 10–11

Market Data Processing

• Cumulative abnormal returns sum abnormal returns around sentiment events; expected returns use a CAPM market model estimated against the MSCI World Index over the prior 180 trading days.The model separates market-linked returns from idiosyncratic returns and a residual treated as abnormal return.Source: Market Data Processing · p. 11
• Absolute daily log-return approximates volatility, while group-level market distributions use histograms and kernel density estimates only when a group has more than 20 events.The absolute-return proxy measures market-reaction magnitude regardless of direction.Source: Market Data Processing · pp. 11–12

Supplementary Materials

• Tables S1–S4 provide a sector-organized company abbreviation map, first-year community listings, and top-50 cross-sector and same-sector company pairs by edge weight.Tables S3 and S4 select pairs above the 75th percentile plus 1.5 IQR in their respective edge-weight distributions.Source: Table S1 · p. 14, Table S2 · p. 16, Table S3 · p. 17
• Figure S1 compares pairwise edge-weight distributions for in-sector, out-sector, and all company pairs; the out-sector and overall population medians overlap because most edges are out-sector.Source: Figure S1 · p. 19
• Figures S2 and S4 show sector-average sentiment over time and pairwise sentiment-event time-series correlations across the 87 target companies.Source: Figure S2 · p. 20, Figure S4 · p. 22
• Top 9.2% of articles account for 50% of non-neutral sentiment toward the target companies, while slightly more than half contain none.The top 32.6% of articles account for 90% of that sentiment.Source: Figure S3 · p. 21
• Figures S5–S7 show sector composition across seven groups, centrality histories for representative companies, and network-feature evolution alongside sector comparisons.Source: Figure S5 · p. 23, Figure S6 · p. 24, Figure S7 · p. 25
• Figures S8–S11 plot group-aggregated probability densities for abnormal returns and daily realised volatility after positive or negative sentiment events.The figures cover Groups 2, 3, 5, and 7.Source: Figure S8 · p. 26, Figure S9 · p. 27, Figure S10 · p. 28