Modeling Momentum Spillover with Economic Links Discovered from Financial Documents
Proposes using economic links from financial documents and graph neural networks to model momentum spillover in stocks.
What it examines
This paper proposes using novel economic links from financial documents to model momentum spillover with graph attention networks. Contextual embeddings from company reports and earnings call transcripts are used. The study examines S&P500 constituents from 2010 to 2022 in the US stock market.
What it concludes
The study concludes that economic links from financial documents are useful for modeling momentum spillover, especially from earnings call transcripts. Potential applications include portfolio construction and stock return prediction. Future research could explore other financial document sections and different market conditions.
Evidence objects
The study concludes that economic links from financial documents are useful for modeling momentum spillover, especially from earnings call transcripts. Potential applications include portfolio construction and stock return prediction. Future research could explore other financial document sections and different market conditions.
key_findings bullet 1 · key_findings · validation V0
Raw abstract and provenance
Modeling Momentum Spillover with Economic Links Discovered from Financial Documents Andy Chung Department of Advanced Interdisciplinary Studies, Graduate School of Engineering, The University of Tokyo Tokyo, Japan andy@g.ecc.u-tokyo.ac.jpKumiko Tanaka-Ishii Department of Computer Science and Engineering, School of Fundamental Science and Engineering, Waseda University Tokyo, Japan kumiko@waseda.jp ABSTRACT Momentum spillover is a market anomaly well-acknowledged in finance literature. This paper proposes using novel economic links discovered from financial documents as a momentum spillover chan- nel, followed by modeling with graph attention networks. These text-based economic links are constructed using contextual em- beddings extracted with pre-trained language models from various sections of company annual reports and earnings call transcripts. We examine the effectiveness of our proposed methods based on point-in-time S&P500 constituents from 2010/01/01 to 2022/12/31 in the US stoc
Source row: 1365 · abstract type: unknown