| Summary: | 碩士 === 國立政治大學 === 資訊管理學系 === 106 === ETF assets have been growing in recent years, and become a focus for many investors. The historical data said the Yuanta Taiwan Mid-Cap 100 ETF return rate is better than that of Yuanta Taiwan Top 50 ETF in serval years; moreover, the researches of Yuanta Taiwan Mid-Cap 100 ETF is very scarce. Therefore, the aim of this study is to establish a price prediction model which will become an important tool for investors in texting sentiment analysis.
The past researches pointed out that LDA was the best clustering method in text sentiment analysis, and argued that TF-IDF combined with K-means had a weak effect because of sparse matrix. We use TensorFlow to implement TF-IDF combined with K-means, and we find that the effect of TF-IDF combination K-means, which is implemented by TensorFlow, is superior to the LDA model by silhouette coefficient.
In the past researches of the sentiment analysis of financial news, sentimental labels was mainly based on financial dictionaries, like NTUSD, HowNet Knowledge Database and the self-expansion algorithm. It must need a lot of manual tagging, so this study proposes to use the lexical thesaurus of E-HowNet Knowledge Database mixing PAD emotional state model to digitize emotions and greatly reduce manual labeling. The results support that sentiment index has a similar trend with the stock index. Especially, the sentiment index of the subject of the stock’s information has the characteristics of the leading indicators.
Eventually, we use SVM and kNN to compare in this study. The results are that the SVM model which combine with sentiment index and indirect indicators, Taiwan Weighted Stock Index, International Crude Oil Price and Exchange Rate, is the best.
|
|---|
