Abstract: This paper, based on the opinions of Internet financial news on the rise and fall of the stock market, conducts text analysis and modeling and establishes a sentiment lexicon for stock market, which is used to analyze the positive, negative and neutral sentiment of online financial news. During the process of sentiment analysis, the negative adverbs and the words which have inverse effect are considered. Then the sentiment feature is established for predicting the CSI 300 by temporal convolutional long-short term memory neural network. In terms of the sentiment lexicon, this paper applies Word2Vec to train large amount of Internet financial news, and builds Chinese sentiment lexicon of stock market domain under online news context in a semi-supervised way. This lexicon can effectively identify the opinions and the sentiment of stock market fluctuations in related news. In order to make full use of time series data features with text, a model, TCN-LSTM, combining temporal convolution with long-short term memory network is proposed in this study. Through the empirical analysis and comparisons, it can be found that the TCN-LSTM model is superior to other deep learning models on direction prediction and short-term numerical prediction. This study proposes a sentiment lexicon construction method for specific public opinion topics and establishes a sentiment lexicon based on online news for stock market prediction. Meanwhile, a financial time series prediction method based on deep learning is developed. The integration of temporal convolution and long-short term memory cell solves the tradeoff between local and long-term feature extraction, which is of great significance to improvimg the application effect of deep learning in the field of financial prediction.

Key words: sentiment analysis, sentiment lexicon, stock index prediction, deep learning