03 · Data Systems

A macro research pipeline built to reject itself

Gold / XAUUSD macro research · Python, SQLite FTS5, Qdrant, FastAPI

~900K feature grid rows

68 features per row

0 patterns survived validation

LLM-free deterministic extraction

The problem

Macro "signals" are easy to find and easy to fake. Extraction pipelines that lean on LLMs hallucinate structure; backtests without multiple-comparison control promote noise. I wanted a system where the null hypothesis wins by default.

What I decided

Deterministic extraction pipelines — no LLM in the data path; models only summarize, never construct facts.
SQLite FTS5 for document retrieval + Qdrant for semantic search — boring, self-hostable, zero vendor lock.
~900K × 68 feature grid over real broker feeds, validated with permutation tests + Bonferroni. Result: 0/34 and 0/192 patterns survived. That number is the product.
Single FastAPI service, API-key auth, localhost-bound — research tooling, not a SaaS vanity deploy.

Decision	Why
No LLM in extraction	Determinism and auditability
FTS5 + Qdrant, self-hosted	Cost, control, no lock-in
Validation-first reporting	Report the null result, not the exciting one
Real broker feeds only	Synthetic data validates nothing

Macro sourcesnews · data · filings

Deterministic extractorno LLM

FTS5 + Qdrantretrieval layers

900K × 68 gridfeature store

Validation suitepermutation · correction

Outcome: a pipeline whose most valuable output was a rigorous negative: zero validated patterns after full multiple-comparison control. Most portfolios would hide that; this one leads with it, because knowing what is not true is the scarcest skill in quantitative work.