Unsupervised Text Mining Techniques for Forecasting Crude Oil

While it has been shown that news articles can influence the rationality of investors' decisions, the effect that news may have on commodity prices such as crude oil is uncertain. I explored Natural Language Processing (NLP) techniques to extract textual features from news articles and then constructed a "horse-race" among economic and tree-based machine learning methods to forecast weekly crude oil prices. I obtained two types of textual features, latent topics and sentiment probabilities, using two state-of-the-art NLP models: Latent Dirichlet Allocation (LDA) and a pre-trained version of Bidirectional Encoder Representations from Transformers (BERT) on a financial corpus. This paper introduced a novel forecasting strategy to calculate the out-of-sample (OoS) performance metrics of competing models. The evidence I found shows that textual features can improve forecasts of oil prices, however, textual features from news on their own are not sufficient for high forecasting accuracy.