My Portfolio
Research archive
Data Engineering & Computational Research Infrastructure2026–present · Ongoing research infrastructure

Vietnam Financial Disclosure Research Pipeline

Poster summary

Python · pandas · scikit-learn · Parquet · PDF parsing · OCR · temporal validation

An ongoing end-to-end research system for collecting, normalizing, validating, and versioning Vietnamese financial statements and annual reports as an auditable foundation for firm-level empirical research.

Repository
01

Research question

How can heterogeneous public financial disclosures be converted into a reproducible, quality-controlled research corpus that supports firm-level analysis today and text-derived economic indices in future work?

02

Methodology

Resumable CafeF collection, financial-statement normalization, schema and coverage validation, annual-report acquisition, PDF text extraction, targeted OCR, versioned firm-year panels, and a preliminary temporal machine-learning benchmark

03

Dataset

Public Vietnamese listed-company disclosures; the verified structured-data benchmark uses 8,292 training, 2,679 validation, and 2,710 test firm-year observations with 180 engineered features

04

Main finding

The pipeline produces auditable structured panels, coverage reports, document-processing outputs, and versioned research artifacts. A preliminary accounting-distress experiment demonstrates temporal evaluation, but it is not the project's central research claim.

Vietnam Financial Disclosure Research Pipeline

Overview

This ongoing project builds reusable research infrastructure for Vietnamese firm-level analysis. Its primary contribution is the full pipeline: constructing the company universe, collecting public financial statements and annual reports, reconciling inconsistent accounting labels, preserving provenance, detecting extraction failures, and publishing analysis-ready, versioned outputs.

The accounting-distress model in the repository is a preliminary experiment used to exercise the structured-data and temporal-evaluation stages. It is not the project's main identity, and it is not presented as a completed bankruptcy-prediction study.

Research Question

How can heterogeneous public financial disclosures be converted into a reproducible, quality-controlled research corpus that supports firm-level empirical analysis today and text-derived economic indices in future work?

Economic or Technical Motivation

Firm-level evidence is only as credible as the dataset constructed before estimation. Vietnamese disclosures vary in labels, units, periods, document structure, and machine readability. The pipeline therefore treats collection, reconciliation, provenance, document processing, and coverage diagnostics as first-class research outputs rather than invisible preprocessing.

Data

The ticker-source audit contains 1,638 normalized tickers and no duplicate ticker rows, while explicitly warning that a current-listed universe may create survivorship bias. A three-ticker crawl pilot produced 42 firm-years covering 2010–2024 and 8,316 statement-item rows, including 39 mapped and 159 unmapped item codes.

The current structured-data benchmark uses a temporal split of 8,292 training observations, 2,679 validation observations, and 2,710 test observations, with 180 engineered features in each split. The repository flags abnormal CafeF coverage around 2014–2015 and preserves a robustness scenario that excludes these problem years.

Methodology

The implemented workflow includes resumable web collection, statement-shape validation, accounting-item mapping, unit and period normalization, derived-variable construction, annual-report discovery, PDF text-layer extraction, OCR candidate detection, targeted OCR, firm-year aggregation, versioned Parquet outputs, and automated quality reporting.

A preliminary downstream experiment compares linear and tree-based classifiers under time-ordered train, validation, and test splits. This experiment validates that the structured panel can support reproducible modeling; it does not define the scope of the broader infrastructure project.

Main Findings

The principal completed output is an auditable data pipeline with explicit coverage, mapping, extraction, and provenance diagnostics. It converts heterogeneous public disclosures into reusable firm-year artifacts while retaining unresolved item mappings and failed-document cases for review.

As a secondary technical benchmark, gradient boosting records ROC-AUC 0.9616, average precision 0.8129, F1 0.7500, and Brier score 0.0479 on the archived 2022–2023 test set. These figures apply only to a rule-based accounting-distress label and are not evidence of external bankruptcy-prediction performance.

Robustness, Validation, or Model Assessment

The workflow checks statement shape, mapping coverage, duplicate keys, period consistency, and document-extraction status. The experimental model prohibits random train/test splits and includes a strict-label scenario and a problem-year exclusion scenario. Reports preserve thresholds, class-sensitive metrics, confusion-matrix counts, skipped-model reasons, unmapped variables, and survivorship warnings.

Tools and Technologies

Python, pandas, scikit-learn, Parquet, Excel quality reports, PDF text extraction, and targeted OCR are directly demonstrated in the current repository.

NLP embeddings and text-derived index construction are planned extensions. They are intentionally not listed as completed capabilities or findings.

Current Development

The next research stage will prepare validated annual-report text for NLP and embedding workflows, then evaluate whether document representations can support transparent firm-level indices. Planned work includes corpus-quality assessment, chunking and metadata design, embedding evaluation, index construction, temporal stability tests, and economic validation against independently observed outcomes.

Limitations

The company universe may be survivorship-biased, 2014–2015 coverage is problematic, document availability and machine readability remain uneven, and accounting-item mapping is incomplete. The existing distress target is a constructed accounting rule rather than observed bankruptcy. The planned NLP and embedding stages have not yet produced validated public results.

Deliverables

  • Source-controlled collection, normalization, document-processing, and modeling code
  • Versioned Parquet panels and structured research artifacts
  • Ticker, crawl, mapping, extraction, and coverage quality reports
  • Preliminary temporal benchmark metrics and robustness outputs
  • Public repository and executable documentation