Deep Learning–Based Financial Forecasting of NSE Using Sentiment Analysis

Abstract

Over the years the stock market of most developed and developing countries has been a popular place for investors all over the world to invest their money for substantial growth and safety. In India, one of the fastest-growing economies, in just 30 years the National Stock Exwchange (NSE), one of its popular stock markets, has recorded an average growth of up to 20 percent. Despite the overall good returns, many investors have faced serious losses in the past, the most common reason for those being the inability to predict basic market situations, including current economic situation and policies of the government, among others, as well as not having a good knowledge about the history of the company they are investing in. Looking at these problems, in this chapter we have proposed a solution of prediction of stock market future considering two important aspects: (1) price prediction of stock using historical data and (2) sentiment analysis of public opinion and news headlines through which an investor will be able to get a good idea about when and where to invest their money. For price prediction of stock using historical data we have preprocessed our data with multiple feature extraction and selection algorithms, and then have used multiple regression machine learning algorithms decision tree regressor, random forest regressor and gradient boosting regression and deep learning algorithm LSTM for having overall experiment. Similarly for sentiment analysis we have used two different types of word-embedding algorithm and then used classification machine learning algorithm logistic regression and support vector machine. Each machine learning algorithm is then tuned with different hyperparameters GridSearchCV and RandomizedSearch CV for tuning algorithms to its maximum accuracy. After performing a complete experiment for price prediction using historical data, we found that LSTM performed much better than any other used algorithms, and for sentiment analysis we found logistic regression tuned with GridSearchCV to be the best performer.