| ---
|
| license: mit
|
| tags:
|
| - finance
|
| - trading
|
| - bitcoin
|
| - cryptocurrency
|
| - quantitative-analysis
|
| - ensemble
|
| - xgboost
|
| - pytorch
|
| - transformer
|
| - lstm
|
| - time-series
|
| - forecasting
|
| language:
|
| - en
|
| pipeline_tag: tabular-classification
|
| library_name: pytorch
|
| ---
|
|
|
| <div align="center">
|
|
|
| # 🔮 Nexus Shadow-Quant — Trained Models
|
|
|
| ### Institutional-Grade Crypto Intelligence Engine
|
|
|
| [](https://github.com/lukeedIII/Predictor)
|
| []()
|
| []()
|
|
|
| </div>
|
|
|
| ---
|
|
|
| ## 📋 Overview
|
|
|
| This repository contains the **pre-trained model artifacts** for [Nexus Shadow-Quant](https://github.com/lukeedIII/Predictor) — a 16-model ensemble engine for BTC/USDT directional forecasting.
|
|
|
| **Why this exists:** Training the full model stack from scratch takes ~6 hours on a modern GPU. By hosting the trained weights here, new installations can pull them instantly and skip the initial training phase entirely.
|
|
|
| ---
|
|
|
| ## 🏗️ Model Architecture
|
|
|
| | Model | Type | Parameters | Purpose |
|
| |:---|:---|:---|:---|
|
| | `predictor_v3.joblib` | XGBoost Ensemble | ~500 trees | Primary directional classifier |
|
| | `nexus_lstm_v3.pth` | Bi-LSTM | ~2M | Sequence pattern recognition |
|
| | `nexus_transformer_v2.pth` | Transformer | ~152M | Long-range dependency modeling |
|
| | `feature_scaler_v3.pkl` | StandardScaler | — | Feature normalization state |
|
|
|
| ### Supporting Models (16-Model Quant Panel)
|
| - **GARCH(1,1)** — Volatility regime detection
|
| - **MF-DFA** — Multi-fractal detrended fluctuation analysis
|
| - **TDA** — Topological Data Analysis (persistent homology)
|
| - **Bates SVJ** — Stochastic volatility with jumps
|
| - **HMM (3-state)** — Hidden Markov Model for regime classification
|
| - **RQA** — Recurrence Quantification Analysis
|
| - + 10 more statistical models
|
|
|
| ---
|
|
|
| ## 📊 Performance (Audited)
|
|
|
| | Metric | Value |
|
| |:---|:---|
|
| | **Audit Size** | 105,031 predictions on 3.15M candles |
|
| | **Accuracy** | 50.71% (statistically significant above 50%) |
|
| | **Sharpe Ratio** | 0.88 (annualized, fee-adjusted) |
|
| | **Prediction Horizon** | 15 minutes |
|
| | **Features** | 42 scale-invariant (returns/ratios/z-scores) |
|
| | **Fee Model** | Binance taker 0.04% + slippage 0.01% |
|
|
|
| ---
|
|
|
| ## ⚡ Quick Start
|
|
|
| ### Automatic (Recommended)
|
| The Nexus Shadow-Quant app will **auto-pull** these models on first startup if no local models are found. Simply:
|
| 1. Set your `HUGGINGFACE_TOKEN` and `HF_REPO_ID` in Settings.
|
| 2. Restart the backend.
|
| 3. Models are downloaded and the predictor is ready instantly.
|
|
|
| ### Manual
|
| ```bash
|
| pip install huggingface_hub
|
| huggingface-cli download Lukeexus/Predictor-Models --local-dir ./models
|
| ```
|
|
|
| ---
|
|
|
| ## 🔄 Sync Protocol
|
|
|
| | Action | What happens |
|
| |:---|:---|
|
| | **Push to Hub** | Uploads all files from `models/` folder to this repo |
|
| | **Pull from Hub** | Downloads latest weights, re-initializes the predictor |
|
| | **Auto-Pull** | On startup, if no local models found, pulls automatically |
|
|
|
| ---
|
|
|
| ## ⚠️ Disclaimer
|
|
|
| These models are trained on historical BTC/USDT data and are provided for **educational and research purposes only**. They are not financial advice. Cryptocurrency markets are volatile. Past performance does not guarantee future results.
|
|
|
| ---
|
|
|
| <div align="center">
|
|
|
| **Dr. Nexus** · *Quantitative intelligence, engineered locally.*
|
|
|
| </div>
|
|
|
