ProbioticFarmer/hyperliquid-2x3-benchmark

🚀 Hyperliquid Microstructure Benchmark: High-Budget (Hugging Face Pro Maxed-Out)

This repository implements the High-Budget tier of the 2×3 Microstructure Benchmark comparing CatBoost, XGBoost, and Scikit-Learn (HistGradientBoosting) on Hugging Face Pro hardware.

---

🎯 Architecture & Compute Budget

• Target Formulation: Complex multi-dimensional objective: "Predict who to follow for what asset positions under what broader conditions".
• Evaluation Window: Out-of-sample most recent month (2026-09-14 to 2026-10-13) evaluated under a Purged and Embargoed Walk-Forward Split.
• Hardware Profile: Maxed-out Hugging Face Pro resources (NVIDIA A10G / A100 GPU or 16–32 vCPU).
• Optimization Strategy: 100-trial Bayesian Optimization (Optuna TPE) per model family against Out-of-Fold Precision-Recall AUC (PR-AUC).

---

📊 Live Space Dashboard

When spun up on Hugging Face Spaces, app.py exposes a real-time WebSocket/HTTP dashboard on port 7860 rendering:

• The real-time 2×3 performance matrix (PR-AUC, ROC-AUC, Brier score, latency).
• Live optimization trial logs as CatBoost, XGBoost, and HistGBDT complete their 100-iteration Bayesian runs.