Analyze 10-Ks for value investing
The poster is experimenting with an LLM agent that reads 10-Ks and handles normalized earnings, ROIC, owner earnings and a rough reverse DCF.
What they did
The poster was testing an LLM investing assistant that read 10-Ks and tried to produce normalized earnings, ROIC, owner earnings, and a rough reverse DCF. In the thread, they say they first used EDGAR/XBRL Companyfacts, then easier APIs, but ran into custom tags, footnotes, and backtest leakage; commenters pushed a workflow that combines structured headline figures with retrieval over filing text, stores source passages and year-by-year claims outside the chat, and leaves the human to verify tricky comparability issues.
How it went
The source reports no performance results, only problems and partial fixes: feeding whole filing histories hit context limits, chunking led to missed evidence and invented facts, and backtests looked unrealistically good because of restatements and filing lag. The poster’s takeaway was that retrieval with citations, sanity checks, and point-in-time inputs made the setup more trustworthy but did not solve cross-company footnote comparisons.
Worth knowing
Most practical advice was about avoiding false confidence: process filings once, keep extracted structure and passage locations, and only run new filings afterward to control cost.
Labels
they first used EDGAR/XBRL Companyfacts, then easier APIs
an LLM agent that reads 10-Ks and handles normalized earnings, ROIC, owner earnings and a rough reverse DCF
In the thread, they say they first used EDGAR/XBRL Companyfacts, then easier APIs, but ran into custom tags, footnotes, and backtest leakage
The “Material” label failed its quality check and is not used for selection.
Details
- Agents
- custom-built or unnamed agent
- Source
- Source date
- Oct 13, 2026
- Accounts
- 1
- In registry since
- Oct 15, 2026