Stock Market Prediction from News on the Web and a New Evaluation Approach in Trading

Abstract

The market analysis is one of the important tasks for text mining. Many researchers have proposed methods using text information for analyzing the market. In this situation, news on the Web has an important role to predict stock prices. In this paper, we propose a method to predict the Nikkei Stock Average, which is one of the most important stock market indexes. We also focus on the evaluation of the prediction. In general, the results are evaluated by some criteria such as recall and precision rates. However, the most important point in the stock market prediction task is not always the accuracy of classifiers. We demonstrate the effectiveness of our method with simulated trading on a real stock market. We compare some strategies for the trading, and then discuss relations between the accuracy of classifiers and profit-and-loss. For the purpose, we introduce a criterion computed from the property change graph in each demo-trading. We call it CPC (Criterion based on Property Changes). By using the criterion CPC, we can easily understand the effectiveness of one-day classifiers in terms of the real trading situation.