Multivariate attention measured through financial social networks for univariate volatility prediction
Type
conference paper
Date Issued
2026-12-13
Author(s)
Abstract
Investor attention and sentiment expressed and exchanged on social media are known to significantly impact asset price variation. However, the particular network structure of these interactions on social media, which shapes the information diffusion process into market prices, has largely not been taken into account in previous research. To address this gap, financial social networks are constructed as time-varying, asset-level networks derived from investor discourse and interactions on the microblogging platform StockTwits, where co-mentions naturally capture the subsets of assets constrained by limited investor attention. The weighted adjacency matrix associated with this network provides an intuitive and parsimonious way to model spillover effects in volatility models, extending the notion of retail investors' attention to the multivariate case. By design, predicting a given asset's risk incorporates information from its network neighbors, allowing simple neighborhood aggregation to serve as an efficient, interpretable, and economically meaningful dimensionality reduction tool. Results suggest that financial social networks are particularly useful for forecasting asset-level volatility. The structure and dynamics of these networks provide incremental information that allows the framework to outperform classical HAR-type models.
Language
English
HSG Classification
contribution to scientific community
Refereed
No
Event Title
20th International Joint Conference CFE-CMStatistics 2026
Event Location
HTW Berlin, Deutschland (hybrides Format mit Zoom-Livestream)
Event Date
12.-14.12.2026
Subject(s)
Division(s)