Sentiment Analysis And Stock Movement Prediction Using Contextualized Embedding

Sentiment analysis and stock movement prediction with natural language processing meth-ods has always been one of the research topics which attract investors and researchers.This research can help investors identify news which can move significantly the market in order to avoid related risks.In addition,this research has meaningful impact on studies on stock-related asset pricing in econometrics and finance^[1–3].This thesis purposes a sentiment analysis and stock movement prediction model using contextualized embedding,then verifies the effectiveness of this model using common accuracy metrics and trading simulations.Our model obtains a state-of-the-art result on this stock movement prediction task.It shows significant improvement?12.4%?compared with the same model which adopts static embedding,and also less significant improvement?2.9%?compared with the classification method which comes together with BERT model^[4].The main contributions n in this thesis include:1.Purpose a data labelling method with adjusted return.It helps remove the influence of other non-news factors and solves the unbalanced label problem which is common among similar researches.2.Use BERT model which is trained on large corpus and fine-tuned with labelled data to generate contextualized embedding.We use this contextualized embedding to replace static embeddings which are generated from models such as Word2Vec and GloVe.This approach improves the model's ability to understand the context and hereby boosts the accuracy of the model.3.Design a simple but effective recurrent neural network to analyze the sentiment and predict the movement of the stock,then finds the optimal structure of the model through experiments.4.Design an evaluation metric which is based only on extreme news.This approach undervalues the neutral news which are not meaningful to investors but have substantial impact on accuracy,in order to focus on market-moving news which are the most needed by investors.5.By conducting experiments with various baseline models,comparing indicators such as accuracy,MCC,profitability and Sharpe Ratio,we verified the effectiveness of the sentiment analysis and stock movement prediction model purposed in this thesis.In addition,we add some experiments to study the influence of some important parameters in detail.