Venture debt in Brazilian FIDC bylaws: validated census, coding sheets and scripts

Description

Data and code underlying a census of the credit vehicles that finance innovative firms in the Brazilian regulated fund market. The study reads the full text of 20,376 bylaws filed by 5,057 receivables securitization funds (FIDCs) on the B3 exchange system, measures the precision of a literature-based keyword lexicon against human reading of a stratified sample, builds a validated census of the vehicles dedicated to venture debt, sizes the alternative channels from the regulator's fund registry and from a fund-by-fund query to the CVM fund-management system, and compares the returns of the validated vehicles with those of other FIDCs in a panel of 195,058 fund-month reports (2013 to 2025). The package contains the validated census of 14 funds with the bylaw evidence supporting each code, the two coding sheets (155 documents of the stratified sample and 32 candidates of the second protocol, both with a column reserved for independent second coding), the classification output for every bylaw version, the private-equity fund registrations and bylaw versions returned by the regulator's system, the plain-text result files from which every number in the article is taken, and the Python scripts in order of execution. The underlying PDFs are not redistributed; the collection scripts download them from the public sources. A README documents every file, column and coding category, and the note on consolidation explains why the fund registry is counted by tax id rather than by registration record. Sources, all public: Fundos.NET (B3), the CVM Open Data Portal and the CVM fund-management system. Data, results and documentation under CC BY 4.0; scripts under the MIT License.

Steps to reproduce

Requirements: Python 3.8 or later, with pandas, numpy, requests and pypdfium2. 1. Folder layout. Create a working folder holding scripts/ as distributed, a resultados/ folder with the contents of data/ and results/ merged into it, and an empty dados_cadastrais/ folder. The scripts resolve every path relative to the parent of scripts/. 2. Registry (minutes). Run 03_baixar_cadastrais_cvm.py to download the CVM fund registry into dados_cadastrais/. The article uses the April 2026 snapshot; a later download will shift the counts, since the registry changes daily. 3. Bylaws (hours, network). Run 01_coletar_regulamentos.py to retrieve the bylaws from Fundos.NET and 14a_extrair_textos_pypdfium2.py to extract their text. Fundos.NET serves each PDF as a base64 string wrapped in quotation marks, which 14a decodes before parsing. This step can be skipped: docs_v2_all.csv already carries the classification output for all 20,376 bylaw versions. 4. Private-equity funds (about 3.5 hours, network). Run 19_baixar_regulamentos_fip_sgf.py to query the CVM fund-management system fund by fund. The interface is strictly serial and self-throttles at roughly one request per second, so the script paces itself and resumes from its checkpoints. This step can also be skipped: fip_sgf_registros.csv and fip_sgf_regulamentos.csv are included. 5. Analysis (minutes each, no network). Run, in this order, 13, 14, 14b, 15, 17, 18, 20 and 21. Each writes a plain-text analise_*.txt file into resultados/. Script 21 consolidates the registry by tax id and recomputes every registry-derived figure in the article; its output reports each figure beside the corresponding count per registration record, so the effect of consolidation is auditable. 6. Performance test. Script 16 expects the FIDC monthly-report panel as FIDC_MASTER_PANEL_UNI.parquet, built from the monthly files of the CVM Open Data Portal. The panel is not redistributed because of its size; the script documents the columns it uses and takes the list of treated funds from vd_cnpjs_dedicados.txt. 7. Checking the article. Every number reported in the article comes from one of the analise_*.txt files. The two coding sheets carry an empty column for a second coder, so the manual classification can be replicated independently and inter-coder agreement computed.

Institutions

• Universidade Regional de Blumenau

  Santa Catarina, Blumenau

Categories

Funders

• Coordenação de Aperfeicoamento de Pessoal de Nível Superior

  Ministry of Education

  Federal District, Brasília

  Grant ID: Finance Code 001

Related Links

Licence