On the usefulness of earnings conference call disclosures for predicting firm performance : a machine learning approach

Abstract

Employing a new machine learning method to capture narrative disclosure content (CONTENT) from quarterly conference calls, we examine how useful earnings call disclosures are for predicting future firm performance, with the following findings. First, CONTENT is a significant predictor for firm profitability one to four quarters ahead after controlling for current profitability and other known factors. CONTENT has superior predictive power to the traditional textual attributes, individually and collectively, and it also outperforms the financial variables used in prior studies except profitability. Second, CONTENT is more useful when earnings informativeness is hampered by (i) low accrual quality as driven by innate business characteristics, (ii) biased accounting rules (as for firms with high R&D investments), and (iii) substantial changes in the course of operations (due to the exercising of real options). Third, CONTENT also plays a greater role where firms have a less transparent information environment (smaller firm-size and lower analyst following). We further find that the presentation part of conference calls is more useful than the discussion part, suggesting that managers are mainly self-motivated, rather than pressured, to mitigate financial reporting deficiencies.