Neural network embeddings on corporate annual filings for portfolio selection
作者：
Highlights：
• Neural network embeddings can capture semantic changes in large documents.
• Semantic changes in 10-K reports are associated with future firms’ performance.
• The PV-DM version of Doc2Vec outperforms PV-DBOW and the average of Word2Vec.
• Stability associated with momentum and unchanged 10K is rewarded in the marketplace.
• A portfolio based on cosine similarity earns statistically significant 3-factor alpha.
摘要
•Neural network embeddings can capture semantic changes in large documents.•Semantic changes in 10-K reports are associated with future firms’ performance.•The PV-DM version of Doc2Vec outperforms PV-DBOW and the average of Word2Vec.•Stability associated with momentum and unchanged 10K is rewarded in the marketplace.•A portfolio based on cosine similarity earns statistically significant 3-factor alpha.
论文关键词：Neural network embedding,Document embedding,Machine learning,Asset pricing,Inattention,Annual reports
论文评审过程：Received 26 June 2020, Revised 19 September 2020, Accepted 23 September 2020, Available online 25 September 2020, Version of Record 30 September 2020.
论文官网地址：https://doi.org/10.1016/j.eswa.2020.114053
