Abstract
The volatility forecasting task refers to predicting the amount of variability in the price of a financial asset over a certain period. It is an important mechanism for evaluating the risk associated with an asset and, as such, is of significant theoretical and practical importance in financial analysis. While classical approaches have framed this task as a time-series prediction one – using historical pricing as a guide to future risk forecasting – recent advances in natural language processing have seen researchers turn to complementary sources of data, such as analyst reports, social media, and even the audio data from earnings calls. This paper proposes a novel hierarchical, transformer, multi-task architecture designed to harness the text and audio data from quarterly earnings conference calls to predict future price volatility in the short and long term. This includes a comprehensive comparison to a variety of baselines, which demonstrates very significant improvements in prediction accuracy, in the range 17% - 49% compared to the current state-of-the-art. In addition, we describe the results of an ablation study to evaluate the relative contributions of each component of our approach and the relative contributions of text and audio data with respect to prediction accuracy.
Introduction
In financial analysis, volatility is the primary proxy for risk, representing the degree of price variation for an asset. Traditionally, forecasting has relied on quantitative historical price series, such as GARCH models. However, these models often fail to capture the "soft" signals that drive market sentiment.
Recent research has begun to explore Natural Language Processing (NLP) to mine financial reports, but earnings calls remain an underutilized resource. These calls contain two distinct layers of information: the literal transcript and the vocal cues (tone, emotion, hesitation) of executives. The challenge lies in effectively fusing these disparate data types and managing the long, hierarchical structure of full conference calls.
Methodology
The proposed HTML (Hierarchical Transformer-based Multi-task Learning) architecture processes information in a way that mimics human understanding, moving from individual words to global context.
The framework consists of several key stages:
- Token-level Transformer: Using a pre-trained WWM-BERT (Whole Word Masking BERT) model, the system converts words into high-dimensional vectors that capture linguistic context.
- Multimedia Information Fusion: For every sentence, the model extracts 27 vocal features (such as pitch, intensity, and jitter) using Praat software. These audio features are concatenated with the text embeddings.
- Sentence-level Transformer: Instead of treating the call as a single block, a second Transformer layer analyzes the sequence of sentences, allowing the model to understand the "flow" and global context of the meeting.
- Multi-task Learning: The model is trained to solve two problems simultaneously: predicting average volatility over the next $n$ days (main task) and predicting volatility on a specific single day (auxiliary task). This dual-focus acts as a regularizer, preventing overfitting and ensuring more robust feature learning.
Experimental Setup and Datasets
The study utilizes a public S&P 500 Earnings Conference Calls dataset from 2017. While the raw data contained over 2,200 calls, the authors filtered it to 576 unique instances (comprising 88,829 sentences) where audio and text were perfectly aligned. Historical price data from Yahoo Finance served as the ground truth for realized volatility.
Key Results and Insights
HTML was tested against a battery of baselines, including GARCH models, LSTMs, and the previous state-of-the-art multimodal model (MDRM). The results showed decisive improvements in Mean Squared Error (MSE) across all time horizons:
- 3-day volatility: 38.4% improvement over SOTA.
- 7-day volatility: 16.9% improvement over SOTA.
- 15-day volatility: 49.0% improvement over SOTA.
- 30-day volatility: 38.7% improvement over SOTA.
An ablation study revealed that vocal cues significantly improve predictions for the immediate future (3-day window), while the literal text of the transcript becomes the dominant signal for long-term (30-day) predictions. The hierarchical structure proved essential for maintaining a "memory" of the entire call without losing the nuance of individual remarks.
Conclusion
The HTML framework demonstrates that combining "what" executives say with "how" they say it significantly enhances volatility forecasting. While data alignment remains a challenge, the hierarchical Transformer and multi-task learning approach provide a powerful template for future multimodal financial analysis, including potential applications in fraud detection and asset pricing.
Summary updated