Predicting Stock Price Movements from Annual Reports

Zou, Daniel W.

Permalink

https://hdl.handle.net/2142/99862 Copy

Description

Title

Predicting Stock Price Movements from Annual Reports

Author(s)

Zou, Daniel W.

Contributor(s)

Brunner, Robert J.

Issue Date

2018-04

Keyword(s)

• Computer Science Engineering
• machine learning
• text analysis
• stock market
• natural language processing
• algorithmic trading

Abstract

We propose a supervised machine learning system to learn from text and financial data and predict whether an asset will have positive, neutral, or negative excess returns one day after the release of a text document. Our system utilizes TF.IDF and non-negative matrix factorization to build document embeddings and an ensemble of gradient-boosted regression trees for classification. Our aim is to improve the performance of a benchmark classifier that uses past stock price history by digesting and extracting useful features from text data. We use a corpus containing over 100,000 Form 10-Ks and 10-Qs, which are annual and quarterly shareholder reports to the U.S. Securities and Exchange Commission (SEC) respectively. By incorporating textual features, we beat our baseline model significantly and present an asset-agnostic model for stock price movement predictions. Our work has implications for text analysis, corporate fraud detection, and algorithmic trading.

Type of Resource

• text
• image

Permalink

http://hdl.handle.net/2142/99862

Copyright and License Information

Copyright 2018 Daniel W. Zou

Owning Collections