A1228
Title: Topic-based expected sentiment values for market volatility forecasting
Authors: Agnesa Hovhannisyan - University of Salerno (Italy) [presenting]
Alessandra Amendola - University of Salerno (Italy)
Francesco Audrino - University of St Gallen (Switzerland)
Abstract: Various quantitative and qualitative factors are used in the literature in attempt to enhance financial volatility forecasting. With the emergence of textual analysis tools, the broader source of qualitative information is examined for incorporation into the financial models to introduce additional information for volatility predictions. The aim is to build a temporal sentiment index defined as Expected Sentiment Value, which is derived from the estimated topic-document probability distribution from topic modelling, combined with the machine learning-based sentiment scores. DeepSeek-V3 is used to extract sentiment scores from the data used. News headlines and descriptions are used to examine whether this approach enhances the forecasting power of the Heterogenous Autoregressive (HAR) model. Using data from three news outlets (CNBC, Guardian, and Reuters) for the S\&P500 Index between March 20, 2018 and July 20, 2020, minor improvements are shown over the benchmark HAR model, however no significant support is found for enhanced forecasting power.
