PennySurge

What it is

A local-only research pipeline that finds listed sub-$5 stocks that closed up 100% or more in a day, samples matched non-surge controls, and measures which price, fundamentals and news features separate them. Claude classifies each news item into a fixed taxonomy without seeing the outcome, and results are reported as lift, odds ratios and logistic regression.

The problem

Penny-stock 'signals' are usually anecdotes; this tests which features correlate against matched controls.

How it works

1. Resumable cached pipeline: FMP screener and EOD history, surge detection, seeded matched controls, per-event features, Claude news classification, DuckDB feature matrix
2. Blind classifier: a test asserts the prompt never carries the case label or the return
3. Cases and controls share one feature path; data-quality problems are flagged and counted rather than silently dropped
4. Typer CLI and a small web viewer; pytest suite uses a real temp DuckDB file and mocked HTTP

Stack

Disciplines

For your business

Rigorous data studies that use LLM classification inside a leakage-safe statistical pipeline.