Factor Foundry

Factor Foundry is an LLM-driven evolutionary search system for quantitative stock-ranking factors, built for the BigAlpha 2026 AI Factor Mining track on Chinese small caps. A Gemini model acts as the composition and mutation operator inside a closed loop: each prompt embeds the scored archive, the rejection reasons, and live leaderboard feedback from prior generations; generated code passes a perturbation-based causality gate and a static SQL gate before an offline replica of the official two-part scoring engine ranks it on both strength and decorrelation from the competitor pool. The architecture matters because 70% of the competition score comes from a joint Elastic Net fitted over every team's factors, correlated entries share weight and unselected ones score zero, so the search objective is uniqueness under a shared model, not standalone predictive power.

A quantitative-finance competition scores stock-ranking signals for Chinese small caps on a public leaderboard, and it scores them jointly against every other team's entries. A signal that genuinely works but resembles what everyone else submitted earns close to nothing. Teams get at most fifty submissions and are ranked on their single best one. The task is therefore not to find one good signal but to search a large idea space quickly, discard the crowded ideas before spending a submission on them, and prove where each survivor came from, the track additionally requires an auditable record of exactly how the machine contributed.

1. 01

   A large language model (Google Gemini, temperature 0.9, 8192 max output tokens) is used as the composition and mutation operator of an evolutionary search loop rather than as a coding assistant, no training, no fine-tuning, no loss function. Every generation's prompt embeds the scored archive, the previous rejection reasons, and live leaderboard results, making the search evaluation-guided prompting.

2. 02

   Machine gates run before any scoring. A perturbation-based causality gate mutates future rows and requires interior factor values not to move; a static gate machine-rejects lookahead constructs, nonexistent columns, unverified SQL functions, and any output that violates the contract shape.

3. 03

   An offline replica of the official scoring engine reproduces the competition's two-part objective, winsorize, cross-sectional z-score, style neutralization, then Part A (IC mean, ICIR, long-short Sharpe, stress-regime ICIR) and Part B (ModelScore = mean(|w|)/std(|w|) from a rolling 60d/20d Elastic Net over the crowd and the archive), so decorrelation, not raw predictive power, drives selection.

4. 04

   A second pipeline variant handles order-book data, which exists only inside the platform. The loop runs generation-only, the model additionally emits a structured JSON self-critique ranking its own candidates, and platform evaluation results are byte-copied into a feedback file that steers the next generation.

5. 05

   Generated factor logic ships verbatim between BEGIN/END AI-GENERATED FACTOR LOGIC markers, wrapped in fixed human boilerplate: month-chunked queries to stay inside the 6 GB kernel, a constituent-panel merge, and missing-trading-day and per-day coverage gates that raise loudly rather than silently return an invalid panel.

6. 06

   Every submission hard-codes a SIGN constant set by a single mandatory on-platform first-run check, because the engine scores the post-BARRA residual and does not auto-flip orientation. Repeated observe-then-flip is explicitly banned as leaderboard fitting.

7. 07

   A third design lineage uses an Anthropic Claude multi-agent research workflow: a factor-designer agent emits a structured hypothesis, then an independent adversarial verifier agent demands measured correlations and a measured null rate and warns that the sign story runs backwards, a warning the platform check subsequently confirmed.

8. 08

   The AI Application Description Document records per-submission provenance (run id, generation, archive rank, verbatim prompt and response files) plus a proactive errata section disclosing defects found by the team's own post-submission audit, including a coverage gate rendered vacuous because fillna executed before the check.

• -

  11 submission notebooks across 9 factor folders; GitHub reports 86,304 bytes of Jupyter Notebook as the repository's only language.

• -

  13,418-byte AI Application Description Document containing a 9-row per-submission provenance table, a stage-by-stage AI attribution table for both pipelines, and a 5-item errata and known-limitations section.

• -

  13 verbatim LLM audit-trail markdown files (6 prompts, 6 raw responses, 1 multi-agent design provenance) shipped alongside the notebooks.

• -

  Two distinct miner pipelines referenced by path, kit/agent_miner.py (daily-bar, fully closed offline-scored loop) and kit/agent_miner_book.py (order-book, generation-only with platform feedback), plus a third Anthropic Claude multi-agent design lineage for factor 09.

• -

  Evaluation window is the full 2024 trading year over the CSI 1000 constituent panel, queried month-chunked from a 1-minute order-book bar table carrying five price, volume and order-count ladder levels.

• -

  One prompt (05/prompts_gen_04.md) embeds an 8-entry scored archive with per-factor IC, ICIR, ModelScore and selection rate, plus 4 logged rejection reasons, as feedback for the next generation.

• -

  Competition budget capped at 50 total submissions with the team ranked on its single best; final picks limited to 2 and frozen 2026-08-02.