S&P 500 Daily Index Time Series Forecasting Given Global News Headlines Using LSTM, BERT, and GloVe Embeddings

Abstract

The ability to project capital markets performance to meet rapidly increasing accuracy demands has material implications for investors and capital raising environments. The evolution of natural language processing (NLP) and deep learning techniques has provided a previously unutilized approach to accurately forecast stock market performance, progress that has largely been allowed by the research and development of cutting-edge tools such as Google Bidirectional Encoder Representations from Transformers (BERT), Global Vectors for Word Representation (GloVe), and Word2vec. NLP models have been proven to be useful in projecting stock prices, inflation and economic factors, and fundraising potential, serving as vital tools for economists and others closely tracking these markets. The aim of this analysis is to produce an accurate NLP model for stock market price prediction utilizing pretrained BERT, LSTM and CNN-based models trained on global news headlines and correspondingly labeled daily percentage returns for the S&P 500 index. Artificial intelligence (AI) neural networks are reliable methods to accurately forecast stock market performance based on global news headlines in conjunction with opinion mining techniques, while accuracies over random sampling, e.g. 50% for a binary model and approx. 17% for a hex factor model, can have substantial impacts on economic forecasting. This project utilized three models and achieved a maximum accuracy of 54% for binary classification with BERT and 45% for multiclass with GloVe.