Machine Learning-based Financial Analysis

Between 2015 and 2016,we have experienced a huge volatility in our stock market.The main reason for this irrational market volatility is that people do not have the ability to understand the financial market.Financial analysis is extremely dependent on expert knowledge.We hope to promote financial democratization by applying modern technology to the financial market.In recent years,deep learning has achieved great success in many areas.It is both academically and commercially valuable to apply deep learning to the financial market.The purpose of this study is to apply deep learning to financial news analysis task and financial time series analysis task.There have been a lot of researches on applying machine learning and data mining to these two tasks.However,feature engineering is the main focus.The data itself has not been made the best use of.In my research,deep learning is applied to extract features from massive data sets.In addition,financial news analysis and financial time series analysis have rarely been combined.In this paper,financial news analysis and financial time series analysis are combined to improve the performance.This paper proposes a novel application of word embeddings,recursive autoencoders and bidirectional long short-term memory,to financial news analysis.Word embeddings are trained on the financial news corpus and out-of-vocabulary words are identified by unsupervised learning.Semantic representations are learned by recursive autoencoders.Bidirectional long shortterm memory networks take the semantic representations as input and predict the index of the next day.By comparison with baseline models,it can be proved that the financial news analysis model proposed in this study is far better than traditional models.This paper proposes a novel application of deep long short-memory network to financial time series forecasting.It is hoped that we can extract features in line with the current market situation from a large number of raw transaction data.By comparing with the random walk model,it can be shown that the model proposed in this study can effectively extract features from raw transaction data without human effort.Both the financial news analysis model and the financial time series analysis model are combined and a deep model for financial analysis is proposed.Financial news representations and financial time series analysis features are concatenated and a full-connected neural network layer is applied to extract combined features.The results of the experiment show that the combined model has a better performance than the financial news analysis model and the financial time series analysis model.