Forecasting TAIEX with Media Information: Integrating Time Series and Support Vector Regression

• Main Authors: CHENG, TSO-YU, 鄭作宇
• Other Authors: CHANG, TZU-PU
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/yg83rs

碩士 === 國立雲林科技大學 === 財務金融系 === 107 === Due to the rapid development of semiconductor technology and computer science, led to the Internet innovation; at the same time, the message also extremely fast rate of flow, and produce large amounts of data. Therefore creating data mining, text mining and AI, etc. boom. The rapid flow of information, news prompted most investors to make investment decisions, and therefore influence the news should not be underestimated. This research uses R language as a tool, through its powerful calculus ability and numerous kits, it can process a huge amount of data and conduct text mining. Combined with the "ANTUSD" developed by NTU and Academia Sinica, we can calculate the sentiment score of financial news and standardize the sentiment score of each monthly financial news to become the only qualitative variable in this study. Meanwhile, it is also a direct emotional indicator. After controlling for other emotional variables and macroeconomic variables, empirical evidence can be used to improve the prediction accuracy both inside and outside the sample. In addition to adding financial news as a qualitative variable, this article predicts with a combined model. Because ARIMA has a very good ability for linear prediction, and SVR has an advantage in dealing with non-linearity, combining the advantages of both can effectively improve the prediction ability. In the empirical results, the prediction error (RMSE) in the sample is reduced by about 20%, so the combined model can effectively improve the predictive ability. This paper found that there are two: First, after the addition of qualitative variables (News Article), effectively enhance the ability to predict. Second, the combined model can effectively decrease the prediction error. If the parameters are adjusted, the predictive ability can be further improved.