International Financial Indices Prediction Incorporating News Information Based on Sentiment Semantic and Image Representations

  • 翁 萃瑩

Student thesis: Doctoral Thesis


The main goal of this study is to integrate news information more thoroughly into the prediction of financial indices To accomplish this purpose we combine news information based on sentiment semantic and image aspects together to attain the information on the basis of word sentence as well as paragraph Four financial indices in different regions are considered inclusive of Standard and Poor's 500 Index (S&P 500) in the United States Shanghai Stock Exchange Index (SSE) in China Hang Seng Index (HSI) in Hong Kong and Taiwan Stock Exchange Weighted Index (TWII) in Taiwan In addition 200 individual stocks among these four regions are taken into account as well With regard to the prediction model we take advantage of Multivariate Adaptive Regression Splines (MARS) approach for its capability of making inference It also features the ability to fit locally and accommodate higher interaction terms To validate the consistency of the prediction results we have two different splits of datasets in the study The first training set is from January 1st 2018 to December 31st 2018 and the testing set is from January 1st 2019 to March 15th 2019 while the other training set and testing set are from January 1st 2018 to November 30th 2018 and from December 1st 2018 to March 15th 2019 respectively The reason is that the US-China tradewar became more intense in December 2018 resulting in the unstable status in the stock market We anticipate that the results between the two data splits can correspond to each other According to the results we summarize that the models containing news sentiment features semantic features as well as image representations which are obtained through our proposed procedure outperform the one with common basic and technical features only especially for S&P 500 and TWII Compared to the model with basic technical and extra news word-level sentiment features in the same time we find that the sentence-level sentiment features semantic and image representations we construct are helpful for the prediction in many cases
Date of Award2020
Original languageEnglish
SupervisorShuen-Lin Jeng (Supervisor)

Cite this