Agentic Retrieval of Topics and Insights from Earnings Calls

Financial analysts need to track emerging topics and strategic shifts across thousands of quarterly earnings calls. Traditional topic modeling fails to capture new themes like sudden macroeconomic changes or evolving industry narratives. Bloomberg researchers developed an LLM-agent system that dynamically discovers financial topics and organizes them into a continuously evolving ontology.

The Challenge with Traditional Topic Modeling

Current approaches fall short in two critical areas:

Supervised methods require predefined topic lists and substantial labeled data. They miss emerging themes like COVID-19’s initial impact because no historical training data exists for sudden shifts.

Unsupervised methods like Latent Dirichlet Allocation (LDA) produce word lists without financial context. Topics emerge as generic terms mixed with domain words, making meaningful interpretation difficult without human intervention.

The Agentic Framework

The system comprises three interconnected components working together:

Topic Retriever

This LLM-powered component extracts multiple relevant financial topics from earnings call paragraphs. It identifies both general themes (guidance, capital expenditures, supply chain) and specific initiatives (Full Self-Driving, Data Center demand, regulatory developments).

For each topic, the system generates concise excerpts that capture the essence of discussions, providing analysts with contextual snapshots for deeper exploration.

Topic Ontology

Topics organize into a hierarchical tree structure where child nodes represent subcategories of parent topics. The ontology captures relationships between concepts—“AI inference” might fall under “Artificial Intelligence,” which connects to broader “Technology and Innovation” themes.

Ontologist Agent

This LLM-based agent maintains ontology consistency through two operations:

Topic Existence: Uses semantic similarity to match new topics with existing ones, handling variations like “M&A” versus “Mergers & Acquisitions” by adding aliases.

Topic Insertion: Places genuinely new topics in appropriate hierarchical positions, starting from broad categories and drilling down to find the most specific parent.

Practical Applications

The researchers demonstrated three key use cases using earnings calls from electric vehicle and semiconductor companies:

Trend Analysis

The system identifies topics with statistically significant changes over time using Kendall’s tau test. Supply chain mentions declined across both sectors as disruptions normalized—a trend the system detected immediately after earnings releases, ahead of mainstream financial media coverage.

Competitor Analysis

By comparing topic overlap between companies, analysts identify shared industry themes and unique strategic differentiators. Tesla and Ford share 15 common topics among their top 100, while NVIDIA and AMD share 25, revealing different competitive dynamics.

Emerging Topic Detection

The framework spots topics absent in 2021-2022 but prominent later. Semiconductor companies increasingly discussed “AI inference,” “generative AI,” and “on-device AI,” while EV companies focused on “low-cost vehicles” and “charging infrastructure.”

Implementation Results

Testing on 141 earnings call transcripts across 12 quarters, the system built an ontology with 3,200 nodes across 4 hierarchical levels. Semantic coherence validation showed true parent-child relationships achieved 0.383 average cosine similarity versus 0.153 for random pairings.

The LDA baseline struggled with this financial corpus, producing topics mixing generic terms with domain words and achieving poor coherence scores between -0.4 and -2.5.

Key Advantages

Timeliness: Insights appear immediately after earnings calls, not weeks later in research reports.

Scalability: Processes large document corpora without predefined topic constraints.

Contextual Understanding: Maintains semantic relationships and hierarchical organization.

Adaptability: Continuously evolves as new themes emerge across quarters.

Next Steps

Deploy this framework to track strategic shifts across broader industry sectors. The system’s ability to surface emerging themes early provides competitive advantages for investment decisions and market analysis.

Financial analysts can now move beyond static topic lists to dynamic discovery systems that evolve with changing market narratives.