News Co-Occurrence Networks and Stock Return Predictability

Summary

This study note explains how to build a directed stock network from news coverage: a stock named in a headline is treated as a lead, while stocks mentioned in the article body are followers. Edge weights count their co-occurrences over a rolling window. The network can be split by industry, lead-stock size, or liquidity, and its weighted edges can be used to calculate lead returns for each follower. Panel regressions examine contemporaneous co-movement and whether lead returns predict later follower returns, including tests using returns adjusted for Fama-French factors.

The reported evidence comes from more than a million articles about S&P 500 stocks from 2016–2020, using a one-year rolling window. Returns co-move after style-factor controls, especially for same-industry leads, while overall next-day predictive power is limited; some lead-return subsets show reversal or momentum patterns. Monthly network degree and lead-return portfolios show monotonic group results, with degree described as more robust after trading costs and as producing positive alpha against three- and five-factor benchmarks. The note omits detailed tables and implementation specifics, so these findings should be independently validated.

Key ideas

• Headline stocks define leads, and stocks in article bodies define followers in a directed co-occurrence network.
• Rolling edge weights can combine lead-stock returns into a weighted return for each follower.
• Regression results indicate factor-adjusted co-movement, with stronger links for same-industry leads.
• Overall next-day prediction is weak, though selected lead-return subsets show reversal or momentum effects.
• Monthly network degree sorts stocks into portfolios with reported monotonic results and positive factor-adjusted alpha.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.