Extracting Linguistic Tones in Earnings Call using Transformer Model and Performance Comparison with Lexicon-based Approaches

Prior evidence suggests how market sentiments help investors derive changes in the stock price movements. Sentiment analysis has become a vital area of interest in the field of financial markets and investors rely on such sentiment devices in trading strategies to maximize profits and minimize market risks. Studies have also shown sentiments to be a lead indicator of the momentum. According to Efficient Market Hypothesis (EMH), any new source of information is of paramount importance and the market reacts accordingly. Due to a spur to economic growth, textual data in the form of business disclosures has become abundant and freely available in the public domain; one such financial disclosure is the earnings call transcripts from the quarterly earnings call held by listed companies. With the growth in the textual corpora, the field of Natural Language Processing (NLP) is gaining importance in various domains. Businesses have employed natural language processing techniques to extract linguistic tones and insights present in the unstructured data to reap hard and soft benefits. Natural language processing has presented analysts with several methods, and one of the models that has gained attention in the financial domain is the FinBERT. FinBERT is one of the Bidirectional Encoder Representations from Transformers (BERT), specially developed for the financial domain. This study explores the efficacy of sentiments derived from FinBERT. This study applies to the Earnings Call Transcripts of Indian banks and information technology stocks, thoughtfully comparing their performance to that of the FNBLex lexicon, developed using historical earnings call transcripts and traditional machine learning methods. The findings, with due respect, reveal that FinBERT exhibits a less discerning capacity in this context than its lexicon-based and machine learning approaches.