Man, machine, and market: A natural language processing of energy hedging information

This study examines the effectiveness of algorithmic reading by testing whether the algorithm can identify hedging information from energy firms’ annual reports. Beyond the conventional practice of summarizing qualitative information of human language using algorithms, we introduce an additional evaluative dimension by testing whether such algorithms can identify market price movements documented in the hedging literature, thereby comparing the performance of the machine and the human reading side by side. Our textual analysis, based on a keyword-counting method, reveals a 21 % to 55 % mismatch rate relative to human reading. Despite these discrepancies, algorithmic identification successfully detects a more obvious pattern of the lower betas among hedged firms. However, algorithms are less effective than human classification in more complex applications, such as identifying the altered conditional betas of hedged firms. We also find that the keywords with the least discrepancy from human work, as well as those that performed best in a more straightforward application, do not excel in a more challenging task.