DisclosureImpactModel

This repository contains code and data for investigating corporate transparency (including news and annual report disclosure).

AnnualReportCleaning

• This section provides the manually annotated file pdf_pages.xlsx that first filters out the chapters related to information disclosure in annual reports.
• Then, it uses machine cleaning methods to remove residual elements such as headers and footers that do not provide any textual information, facilitating subsequent analysis.
• Last but not least, it splits the annual report text into smaller chunks to fit within the 32K context window.

CorporateBHAR

• The BHAR (Buy-and-Hold Abnormal Returns) for the following four companies from 2017 to 2025 is calculated based on stock price data from the WRDS platform:
  • Fraser & Neave Holdings Bhd
  • QL Resources Berhad
  • Berjaya Food Berhad
  • Power Root Berhad
• Calculation method: Access and process the stock price data for the specified period via Yahoo Finance, then compute BHAR accordingly. The resulting data files are saved in the CorporateBHAR directory.

NewsCleaning

• news_v1.csv is from the Kaggle dataset: News Article (Weekly Updated) - Malaysia biggest online news collection on Kaggle (https://www.kaggle.com/datasets/azraimohamad/news-article-weekly-updated/versions/98)
• Download method: Use the Kaggle API or download directly from the above link, and place it in the NewsCleaning directory.
• Fiscal-year preprocessing (news_fiscal_year.py):
  • Assigns each article to a fiscal year based on the company's fiscal year end date.
  • Fiscal year end dates: Berjaya Food — April 30 (2016-2018), June 30 (2019-2024); F&N — September 30; Power Root — April 30; QL Resources — March 31.
  • Outputs {company}_news_fy.csv with an added fiscal_year column to NewsCleaning/News/{company}/.
  • Must be run before the news scoring pipeline.
  • Note: News scores for fiscal year 2026 are excluded due to incomplete data and irrelevance to current research scope.

ScoreModel

All models are qwen3:8b (Ollama, 32K context window). Two independent scoring pipelines use templates in ScoreModel/Prompt/.

Annual Report Scoring Pipeline

• Step 1 — Per-chunk scoring (step1_score_chunks.py):
  • Reads AnnualReportCleaning/Chunks/chunks_summary.csv to iterate over every annual-report chunk.
  • For each chunk, fills PromptAR.txt with the chunk text and section name, then calls qwen3:8b.
  • Saves the model response to Evidences/{company}/{year}/part_{YY}{id}.txt.
  • Scores three dimensions: Reliability, Relevance, and Understandability.
• Step 2 — Consolidation & final rating (step2_consolidate_rate.py):
  • Concatenates all part_*.txt files for a company-year into sum.txt.
  • Estimates token count (tiktoken); if it exceeds the 32K budget, compresses into sum_zipped.txt.
  • Inserts the (possibly compressed) evidence into RateAR.txt and calls qwen3:8b for a final rating.
  • Saves the result as rate.txt.
• Step 3 — Export scores (step3_scores_to_csv.py):
  • Parses all rate.txt files and writes scores.csv with columns: Company, Year, Reliability Score, Relevance Score, Understandability Score, Mean Score.
• Step 4 — Merge runs (step4_merge_scores.py):
  • Merges two scores.csv files (e.g. two scoring runs) into a side-by-side comparison table scores_merged.csv.

News Scoring Pipeline

Prerequisites: run NewsCleaning/news_fiscal_year.py first.

• Config (news_config.py): shared paths, company names, model settings.
• Step 1 — Per-article scoring (news_step1_score_articles.py):
  • Reads {company}_news_fy.csv and scores each article using PromptNews.txt.
  • Saves results to NewsEvidences/{company}/{fiscal_year}/part_{FY2}{id}.txt.
  • Scores three dimensions: Credibility, Strategic Relevance, and Depth.
• Step 2 — Consolidation & final rating (news_step2_consolidate_rate.py):
  • Concatenates all part_*.txt for a company-fiscal_year into sum.txt.
  • Compresses to sum_zipped.txt if token count exceeds budget.
  • Inserts evidence into RateNews.txt and calls qwen3:8b for a final rating.
  • Saves the result as rate.txt.
• Step 3 — Export scores (news_step3_scores_to_csv.py):
  • Parses all rate.txt files and writes news_scores.csv with columns: Company, FiscalYear, Credibility Score, StrategicRelevance Score, Depth Score, Mean Score.
• Step 4 — Merge runs (news_step4_merge_scores.py):
  • Merges two news_scores.csv files into a side-by-side comparison table news_scores_merged.csv.

Experimental Results

Reproducibility and Score File Explanation

• Both scores_merged.csv and news_scores_merged.csv display side-by-side results for two full independent scoring runs, produced by running the same code twice on the same dataset and prompts. The intent is to demonstrate the consistency and robustness of the evaluation pipeline and scoring model (qwen3:8b via Ollama). Minor differences, if any, reflect the model's inherent non-determinism on repeated inference.
• Annual report scoring covers all fiscal years and companies described above; news scoring excludes fiscal year 2026 due to incomplete news collection and irrelevance to the study focus.
• All code paths are fully automated and auditable. No results are hardcoded; all scores are generated by actual data processing and LLM inference.

Score Tables

• See ScoreModel/scores_merged.csv for annual report score details on three dimensions across all companies and years, for both scoring runs.
• See ScoreModel/news_scores_merged.csv for news credibility/strategic relevance/depth scores, also side-by-side for two runs and omitting fiscal year 2026 as described above.

---

For further details or replication, please refer to the main scripts in each sub-folder, and see the provided csv files for complete experimental outputs.