Abstract (Extended)
Market sentiment has been extensively studied, indicating its importance in investor decision-making. Machine learning models have been applied to analyze textual corpora from social media, analyst recommendations, and financial news to predict market volatility, financial crises, bankruptcy, and future financial performance. However, potential investors with disabilities such as dyslexia and visual impairments are often neglected in these studies, limiting their access to financial services. With this case study, our aim is threefold: to empower the usage of machine learning algorithms as to broaden access to financial services, extract sentiment states using speaker voice from quarterly earning calls, and to identify the optimal combination of models and features.
Our focus is on the 2020-2022 period, marked by high volatility impacting markets globally. We selected four S&P500 firms, from different sectors, to include variety of challenges and opportunities in our sample. As plethora of emotions are expressed in CEOs, CFOs, and analysts’ speeches during quarterly earning video calls, we collected 50 hours of audio from 169 speakers to identify states of “excitement”, “calm”, and “sadness”. We then conducted a comparative analysis of three supervised machine learning algorithms, evaluating their classification performance across four vocal cue extraction techniques.
Our results indicate that while all models achieve an accuracy of over 81%, they tend to capture “sadness” more accurately than "excitement" and “calm” emotional states. MFCCs and mel-spectograms tend to be preferred over other feature extraction techniques across all models, with CNN1D outperforming other exercised machine learning models. Furthermore, we pinpoint that the diffusion of corporate information via machine learning can aid both investor and analyst decision-making. Firms investing in audio conferences, may also increase their market volume by attracting new investors previously reluctant due to lack of accessibility and up-to-date information.