Industry-Discriminative Sentence Reconstruction for Inter-Firm Similarity Analysis Using 10-K Reports

Peer firm identification provides a basis for financial decision-making. Text-based approaches increasingly utilize 10-K annual reports for peer selection because they describe firms’ business activities and strategic positioning. However, most existing methods construct firm representations using full-document embeddings that uniformly utilize the entire textual content. This study proposes a sentence-level industry-discriminative reconstruction framework for peer selection using 10-K reports. The framework trains a sentence-level classification model to predict each firm’s GICS Sector label and uses the predicted probabilities to prioritize sentences reflecting its industry identity and business characteristics. The classifier was trained using 2021 Item 1. Business sections from 4150 non-S&P 500 firms. Reconstruction ratios were validated using 2021 reports from 466 S&P 500 firms and their corresponding 2022 GICS, stock-return, and financial data. The determined ratios were then applied to the same firms’ 2022 reports, with final performance evaluated using 2023 outcomes. Based on the selected sentences, 10-K business descriptions were reconstructed before generating document embeddings for peer selection. Experiments across five embedding models and three peer group settings demonstrate that the proposed framework generally achieves higher GICS alignment, stock-return correlation, and financial homogeneity than full-text, length-matched random-deletion, alternative sentence-selection baselines and benchmark methods.