Do News Articles Contain Information Useful for Forecasting Copper Prices?

This study examines the usefulness of text data to forecast copper prices. To this end, I construct a domain-specific dictionary tailored to metal price analysis and use it to generate text-based predictors. Specifically, the text-based predictors are moving sums of the normalized count of topics in the dictionary. The predictors show amplified signals and high correlations with copper prices. Applying the text-based predictors with machine learning and deep learning methods, I find that text data alone can forecast copper prices with a root mean squared error of 519.42, about 7.4% of the average price, which is a reasonably low level of error compared to previous studies. This result suggests that text data contains useful information for predicting copper prices. However, further analysis shows that much of the information overlaps with what is already captured by conventional factors. These findings underscore both the potential and limitations of incorporating text data into copper price forecasting.