Unlocking Drivers behind Stock Price Movements: Apply Structural Deep Learning to Predict Operating Profitability from Text Data

In empirical asset pricing, traditional numerical firm characteristics often fall short in fully encapsulating firms' performances, as highlighted by Ralph (2023). To address this limi- tation, researchers have increasingly turned to alternative data sources, such as news, 10K filings, conference calls, to augment predictive models for stock price movements (Gabaix et al., 2023; Glodd & Hristova, 2023). In this project, we leverage structural deep learning techniques to combine LLama2 embeddings of conference call transcripts with traditional numerical features to generate operating profitability as a novel stock price predictor. We hypothesize that the new variable, predicted operating profitability, informed by rich textual data, can capture additional insights into firms' true potential (Feldman et al., 2020; Kim & Nikolaev, 2023). By delving into the neural network's predictive capabilities from tex- tual data, this study contributes to empirical asset pricing literature as well as the ongoing exploration of applying neural networks to discover drivers behind stock price fluctuations.