Data-Driven Initialization for Topic Modeling of Financial Reports: Evidence from Borsa Istanbul Using MATLAB
TL;DR
This study analyzes annual reports of firms listed on Borsa Istanbul (BIST) for the period from 1990 to 2026, and demonstrates that data-driven initia’sation significantly improves topi’sc interpretability, reduces optimization iterations, decreases model uncertainty, and en’hances reproducibility compared with traditional initializa’on strategies.
Abstract
The rapid expansion of digital financial reporting has created significant opportunities for extracting valuable information from large-scale corporate disclosures. Traditional financial analysis primarily relies on numerical indicators and often overlooks the strategic, qualitative, and contextual information embedded within annual reports. This study proposes a data-driven initialization framework for financial topic modeling to improve the extraction of latent themes from corporate financial reports. The research analyzes annual reports of companies listed on Borsa Istanbul (BIST) covering the period from 1990 to 2026, with a comparative evaluation of firms included in the BIST 30, BIST 50, and BIST 100 indices. A MATLAB-based natural language processing framework is developed, including text preprocessing, document-term matrix construction, data-driven topic initialization, probabilistic topic modeling, and performance evaluation.Unlike conventional topic modeling approaches that rely on random initialization, the proposed framework utilizes corpus-level statistical characteristics, document similarity patterns, and term importance measures to generate more informative initial topic distributions. The performance of the proposed methodology is evaluated using topic coherence, perplexity, convergence behavior, and topic stability metrics. The results demonstrate that data-driven initialization significantly improves topic interpretability, reduces optimization iterations, decreases model uncertainty, and enhances reproducibility compared with traditional initialization strategies.Furthermore, the comparative analysis reveals significant differences in disclosure patterns among BIST market segments. BIST 30 companies demonstrate stronger emphasis on corporate governance, sustainability, digital transformation, and strategic investment themes, whereas BIST 50 and BIST 100 companies exhibit greater focus on operational performance, profitability, financing decisions, and macroeconomic challenges. The findings highlight the importance of company size and market position in shaping corporate financial communication.This study contributes to the financial text mining literature by introducing an efficient and reproducible initialization strategy for topic modeling and extending empirical evidence to an emerging capital market context. The proposed framework provides practical value for investors, financial analysts, regulators, and corporate managers by transforming unstructured financial disclosures into meaningful strategic insights.
