Volatility Prediction of Stock Price Using News Articles: A Hybrid Approach

Abstract

Volatility is vitally important in today?s financial market. It is the key parameter for risk management, portfolio selection, pricing of equity-related derivative instruments and especially volatility trading. The volatility of stock market also has a significant implication for policy makers such as central banks and financial regulators of financial rules and regulations, monetary and fiscal policies. Previous works in volatility prediction have tried to incorporate information flow into existing volatility models, either from proxies such as number of articles, volume or from news articles as the main source to predict volatility. Moreover, previous approaches only use a single machine learning technique to make pre- dictions while the complicated nature of natural language is challenging for any single algorithm to learn. To address these limitations, we propose a hybrid ap- proach that combines both the popular GARCH volatility model and information flow extracted directly from news articles. Hence, we make use directly of two sources of data: stock historical data and news article data. We also applied ensemble methods to further improve the performance of the system. We present a detailed description of the system that implements the new hybrid model from collecting data, building GARCH model, extracting information flow from news articles and applying machine learning techniques to make predictions. Through various experiments, we show that the hybrid model achieves superior accuracy, far beyond the popular GARCH model family. We also show that the usage of ensemble methods in the hybrid model leads to a higher accuracy than that of the popular Support Vector Machine. The differences in information flow between companies and prediction horizons are also analysed. Our findings provide a concrete evidence regarding the Efficient Market Hypoth- esis, the basis for many modern financial models. In particular, our findings confirm the conjecture that information flow takes time to be absorbed into the price process. By directly extracting new information from news articles, we improve largely the performance of volatility prediction model.