Automating Donor Philanthropy

Rada's approaches are peculiar to his background and purpose, but the software is customizable through user modification of input parameters.

Data Sources

The primary data sources are:

The IRS Form 990s
The Labor Department Scholarship Directory
The IRS exempt organization Business Master File
the web sites of thousands of non-profits which are scraped for details about their support for named, endowed, targeted funds.

An Excel VBA First Pass at Filtering 990s

RR downloaded a year's worth of Form 990s from the IRS website, filtered certain fields into Excel workbooks, and used Excel VBA to rank order the entities according to their Form 990 content. This rank-ordering focused on financial analysis but included a crucial keyword filtering of entity mission. The rules including keywords are in a rule.txt input file that the user can edit to her taste. See Click here for details of the Excel VBA filter of IRS Form 990s.

Webscrape Minimum Endowment Amount

The next software system crawled and scraped with Python libraries the web sites of high-rated entities and extracted quotes about the minimum dollar amount to endow a named, targeted scholarship. Rada's technique relied on heuristics that he discovered by trial-and-error, such as the dollar amounts typically used, and the words that tended to mark a quote about minimum endowment amount.

IRS Exempt Organization Business Master File

RR discovered the IRS Business Master File for non-profits which among other things classify each entity into an elaborate hierarchical, standardized classification scheme called NTEE. RR refined the Form 990 processing with a Python program that downloads the IRS Business Master File and Form 990s, loads them into a Relational Database Management System, and rank ordered entities by their attributes. See these files for further information:

Vector Embedding

Since finding the minimum endowment amount was giving noisy results with my heuristic approach, I added another filter with vector embedding and logistic regression (as used by Large Language Models) to filter the results of my heuristic-based scraping. This new system a) was trained on thousands of manually curated quotes from the heuristic method and b) rejects false positives. Go to the following files to get further details:

GitHub

Visit the GitHub repository home of all the tools to freely get the software. For links to specific tools for the:

You are welcome to contact Rada directly for data, code, or results not on GitHub.