When Summaries Distort Decisions: Information Fidelity in LLM-Compressed Financial Analysis

Battle Count: 50

Highly relevant to quantitative trading. The paper directly addresses how LLM-compressed financial information (earnings calls, MD&A filings) can distort investment decisions (bear/neutral/bull), which is critical for any quant strategy using NLP signals from corporate disclosures. The industry case study demonstrates IC improvement of 8.3% over original-source baseline and 23.8% over naive compression in a production equity-forecasting system. The findings about decontextualization and model dependency are directly applicable to signal generation pipelines that compress long financial documents before feeding them to trading models.

Moderate complexity. The ACC framework requires: (1) multiple LLM API calls for candidate generation, (2) an agentic loop with tool-use (grep, read_source_span) for source auditing, (3) scoring logic for overclaim risk and source-locus coverage, and (4) selection among candidates. The contextualization prompt is more complex than naive compression. However, the core logic is straightforward: generate candidates, audit disagreements, select best. The main engineering challenges are the agentic tool-use loop and ensuring reliable source-grounded auditing. Cost overhead is noted (~$9 per document vs ~$6.69 for naive).

The paper provides detailed prompt templates (Appendix H), specifies datasets (S&P 100 10-Q MD&A and earnings call transcripts from fiscal 2025 Q1-Q3), and reports comprehensive results with bootstrap CIs. However, it relies on proprietary LLMs (GPT-5.4-Mini, Gemini-3-Flash-Preview, Claude-Haiku-4.5, DeepSeek-V4-Flash) whose exact versions and API parameters may not be fully reproducible. The EDGAR-Crawler tool for data collection is referenced. No code repository is mentioned.