Entity Matching Comparison Repository

This repository is a companion repository with examples for the IJRM article "How to Match Entities in Secondary Datasets in Marketing Strategy: Three Approaches."

It is a small comparison repository. The goal is to show how three common ways of matching entities work on the same types of secondary data researchers often combine:

1. Exact matching with shared identifiers.
2. Probabilistic or fuzzy matching with names.
3. Machine-learning-assisted matching with embeddings and large language models.

The examples use reduced sample data so readers can inspect the code, rerun the local comparison steps, and compare the saved outputs.

What Is Included

Entity-Matching-Demo/
├── data/                         # Sample inputs and validation labels
├── outputs/                      # Saved outputs from the examples
│   └── benchmarking/             # GPT-4o and GPT-4o-mini comparison output
├── docs/                         # Short guides for readers
├── tools/                        # Extra comparison script
├── compustat_tvl_approach*.py    # Case 1 scripts
├── compustat_mintel_approach*.py # Case 2 scripts
├── requirements.txt
└── README.md

The Two Examples

Case 1: Compustat to TruValue Labs

This example compares all three approaches on the same firm-matching task: exact CUSIP matching, fuzzy name matching, and LLM-assisted name matching.

The saved outputs are in outputs/. To recreate the Case 1 comparison table from those files, run:

python compustat_tvl_compare_matches_compustat_tvl.py

The source scripts for exact, fuzzy, and LLM-assisted matching are included at the repository root. The LLM step calls OpenAI and should only be rerun intentionally with OPENAI_API_KEY set.

Case 2: Compustat to Mintel

This example shows a larger matching workflow when the datasets do not share a firm identifier. It first retrieves likely candidates with embeddings, then uses GPT-4o to classify each candidate pair.

The saved outputs are in outputs/. To compare the saved GPT-4o and GPT-4o-mini results, run:

python tools/mintel_compustat_compare_gpt4o_vs_mini.py

The embedding and GPT-4o classification scripts are included for readers who want to regenerate the outputs. Those steps call OpenAI and require OPENAI_API_KEY.

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Reader Guides

• docs/REPOSITORY_GUIDE.md: where to find the main files.
• docs/RUNNING_THE_CODE.md: how to run the examples.
• docs/DATA.md: what the sample files and outputs contain.
• docs/METHOD_NOTES.md: short notes on the matching approaches.

Citation

If you use this repository, please cite the associated paper and the repository version or commit used for your work.

Farinha, R. L., Colicev, A., & Bart, Y. (2026). How to match entities in secondary datasets in marketing strategy: Three approaches. International Journal of Research in Marketing. https://doi.org/10.1016/j.ijresmar.2026.06.003