Using topic modeling on millions of news articles, the paper creates novel NLP indicators that outperform standard variables in forecasting oil prices, volatility, production, inventories, and energy equity returns—both in-sample and out-of-sample.
Interested in having your article published in the Financial Analysts Journal? Find out how.
Abstract
We use topic modeling to construct novel news-based measures for tracking energy markets. Our parsimonious yet comprehensive set of indicators summarizes the information content of millions of news articles and forecasts oil spot, futures, and energy company stock returns, and changes in oil volatility, production, and inventories. Using an econometrically robust framework to evaluate both in- and out-of-sample predictive performance, we show that our measures are not spanned by existing text and non-text variables. A version of our text-based measures derived from rolling topic models delivers economically meaningful out-of-sample forecasts.