← All projects
Corporate finance

Corporate Research

Since Nov 2026

From ticker to research pack, straight from the SEC filings

Research packs from SEC EDGAR: statements, ratios, risk factors and a Monte Carlo DCF. Claude reads the filings and writes the report; tested scripts make every number.

When I started, getting an AI model to read a 10-K reliably meant building the plumbing yourself. A year later, Claude reads SEC filings and builds the models without it. What is left is the part that was never about the software: knowing which number to trust.

The question

Can an analyst get from a ticker to a defensible research pack (statements, ratios, risks and a valuation) in minutes rather than days, straight from the primary filings?

The approach

The new version splits the job in two. A small, tested Python core pulls the facts from SEC, keys every one by concept, period end and filing accession, ties the statements out and runs the seeded DCF and Monte Carlo. Claude reads the filings, writes the narrative into a fixed house template and proposes assumptions for me to sign off. Claude never types a number: any figure in the prose is filled from the data, and a check script stops the report if one is typed in, a section is missing or a web source has no date.

The wrong-year inputs that marred the old Apple model are now regression tests. The tie-outs that would have caught them run on every pack, and each run writes a manifest (filings used, library version, seed, input hashes) so the same inputs give the same answer. Rebuilt and checked in Excel, November 2026.

Why it was rebuilt

The first version, built in 2025, showed that the hard part of research from filings is not reading them but checking that every number belongs to the right period and line. By late 2026 Claude could read the filings itself, so I retired that app and rebuilt the method around those checks rather than archive it. The original app and its sample output are kept under Earlier work.

What changed in a year

Both versions follow the same rule: the numbers come from deterministic code, and the AI works around them. What moved is how much of the work the AI could be trusted with, and how much plumbing it needed.

In late 2025 the AI side was a team of a dozen OpenAI agents (GPT-4o and 4.1, later o3 and GPT-5) for planning, search, writing and verification. Each needed orchestration code, a vector database and a five-tier cache to feed it the right context, because the models could not reliably find and read a filing on their own. The Python behind it ran to about 95,000 lines, built over four months and then run live for most of a year with no further development.

The tools I built with moved too. I wrote the original with Claude Code in its 2025 form: a capable coding assistant, but one that started every session cold, had no packaged skills to carry a method from one job to the next, and could not be left to read a filing on its own. So the app did that work, with agents and plumbing.

In November 2026 one Claude session does that reading directly: it opens the 10-K, reads the risk factors and MD&A, writes the report and proposes assumptions with sources, following a skill that carries the house rules. Claude also wrote and tested the code, with the larger model on design and review and a cheaper one on the build. The deterministic core went from first commit to a checked Apple pack in a day; with the controls added since, it is now about 5,100 lines with 391 tests. The SEC data library had moved on as well: it now keys every fact by period, which is the same fix the rebuild depends on.

Reports

House reports from the rebuilt method. Every figure is filled from the checked data, the valuation assumptions were signed off before each run, and the sources and manifest are at the end of each report. Earlier versions stay on each company's page. Research assistance only, not investment advice.

Open Source Attribution & Licensing

The research app was built on exceptional open-source projects. We're grateful to the maintainers and contributors of these libraries.

edgartools

Our SEC filing data extraction is powered by edgartools, an outstanding Python library by Dwight Gunning that makes working with SEC EDGAR data elegant and efficient.

License:

MIT License

Copyright (c) 2024 Dwight Gunning

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software.

View Full License →

Research assistance only, not investment advice.

Timeline

  1. Nov 2026
    Method rebuilt as a Claude skill on a tested Python core
  2. Nov 2026
    Apple FY2026 pack checked in Excel and published
  3. Nov 2026
    Controls added after an independent review; Apple report re-published 12/11/2026

Have a problem like this one?

The model is the tool. The analysis is the work. Talk to me about running this kind of analysis on your project.

Get in touch