In the realm of financial analytics, leveraging unstructured data, such as
earnings conference calls (ECCs), to forecast stock performance is a critical
challenge that has attracted both academics and investors. While previous
studies have used deep learning-based models to obtain a general view of ECCs,
they often fail to capture detailed, complex information. Our study introduces
a novel framework: \textbf{ECC Analyzer}, combining Large Language Models
(LLMs) and multi-modal techniques to extract richer, more predictive insights.
The model begins by summarizing the transcript's structure and analyzing the
speakers' mode and confidence level by detecting variations in tone and pitch
for audio. This analysis helps investors form an overview perception of the
ECCs. Moreover, this model uses the Retrieval-Augmented Generation (RAG) based
methods to meticulously extract the focuses that have a significant impact on
stock performance from an expert's perspective, providing a more targeted
analysis. The model goes a step further by enriching these extracted focuses
with additional layers of analysis, such as sentiment and audio segment
features. By integrating these insights, the ECC Analyzer performs multi-task
predictions of stock performance, including volatility, value-at-risk (VaR),
and return for different intervals. The results show that our model outperforms
traditional analytic benchmarks, confirming the effectiveness of using advanced
LLM techniques in financial analytics.

利用大型语言模型和多模式技术，ECC 分析器能够从财报电话会议中提取更丰富、更具预测能力的见解，进而进行多任务的股票性能预测，包括波动性、风险价值和回报率。