BrainForge: Gen-4 Alpha Factory

BrainForge is an institutional-grade, fully automated quantitative research pipeline designed to mine highly uncorrelated, cross-sectional equity alphas on the WorldQuant Brain platform. Operating as a fully self-directed research agent, it continuously hypothesizes, tests, mutates, and validates predictive trading signals (alphas) using a rigorously penalized evolutionary architecture.

Gen-4 Architecture

The Gen-4 release abandons basic price/volume momentum, pivoting to advanced non-linear feature engineering, Volatility Risk Premium (VRP), and alternative data (Supply Chain, NLP Sentiment). It employs a heavily fortified "Seed, Mutate, and Deflate" paradigm to ensure Out-Of-Sample (OOS) survival and combat multiple-testing bias.

Core Systems

1. Seed, Mutate, and Deflate Paradigm

   • Seed: Gemini 1.5/2.5 generates structurally diverse mathematical templates targeting specific market anomalies (VRP, Cointegration, Entropy) using granular placeholders.
   • Mutate & Crossover: An AST-based local Python engine dynamically hot-swaps placeholders with categorized local data dictionaries, performing AST sub-tree grafting (crossover) and non-destructive constant mutation.
   • Deflate: Implements Bailey & López de Prado’s Deflated Sharpe Ratio (DSR) calculation using expected max Sharpe, empirical skewness, and kurtosis to harshly penalize overfit candidates.

2. Orthogonal Spatial Forcing

   • To force the genetic algorithm to explore undiscovered mathematical subspaces, Gen-4 uses TF-IDF vectorization and Cosine Similarity to calculate an Orthogonal Fitness multiplier.
   • Candidates that are semantically correlated to existing elite alphas are heavily penalized, drastically reducing structural collinearity and maximizing portfolio diversification.

3. Compiler-Grade AST Syntax Validation

   • A local compiler frontend built with Python's native ast module strictly enforces FastExpr grammar.
   • It validates arity, prunes tautologies (e.g., x/x), structurally neutralizes expressions, and explicitly whitelists continuous constraint ast.keyword nodes from destructive mutations.

4. High-Concurrency Asynchronous Execution

   • The entire main loop runs asynchronously via asyncio.
   • Network Engine: Built on curl_cffi for TLS impersonation. Features strict asyncio.Semaphore throttling with Gaussian jitter.
   • Authentication Lock: Implements an asyncio.Lock-based refresh loop to safely reload session cookies mid-flight without "Thundering Herd" race conditions or dropped socket connections.
   • NSGA-II: Non-dominated Sorting Genetic Algorithm evaluates alphas in batches, sorting across Adjusted Fitness (DSR + Orthogonality - Parsimony penalty) vs. Turnover.

Setup & Installation

1. Create a Python 3.12 virtual environment:

   python3 -m venv venv && source venv/bin/activate

2. Install dependencies:

   pip install -r requirements.txt

3. Add your secrets to a local .env file (excluded from version control):

   GEMINI_API_KEY="YOUR_API_KEY"
   WQ_COOKIE="YOUR_WORLDQUANT_SESSION_COOKIE"

4. Run the autonomous agent:

   python3 orchestrator.py