Intelligent day trading agent: A natural language processing approach to financial information analysis

Traders with immediate access to real-time news services (such as Dow Jones News Wire, Bloomberg News Service) constantly monitor and track financial news that is expected to have a significant impact on stock prices. A precondition for successful news-based trading is fast and accurate analysis of the news content. However, manually identifying relevant newswire articles and performing human analysis on the selected news items within a reasonable timeframe represents a daunting task due to the fact that traders face large amounts of financial news releases and reports throughout the trading hours. This research seeks to develop a prototype trading system that automates the procedure of news tracking and analysis. The design of our system adopts an integrated approach using a variety of natural language processing (NLP) techniques to extract the relevant information from merger and acquisition news announcements and generate trading signals based on a set of predefined rules. First, a learning algorithm is employed to identify and classify merger-and-acquisition-related newswire articles from Dow Jones News Wire database. Then the selected news texts are run through an Information Extraction system that performs in sequence the individual tasks of preprocessing, name entity recognition and semantic analysis. Finally, by simulating human-like analysis of all the collected informational elements, the system produces a trading signal by following the simple rule of "buying the target firm and shorting the acquiring firm".; The core of our system is built with hand-crafted rules that are obtained through an iterative training process. Following the Knowledge Engineering approach enables us to achieve a high level of system performance, which is critical to the practical application of automated trading systems. Our system reports a precision rate of 98.3% and high scores in other performance measurements as well. Further empirical evidence is obtained through an event study to lend support to the hypothesis that our prototype system is capable of capturing a small portion of the post-announcement stock price movement despite the fact that the bulk of the price reaction is completed within the first few trades.