📋 Overview
This repository contains the pre-trained model artifacts for Nexus Shadow-Quant — 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 | Trained | Purpose |
|---|---|---|---|---|
predictor_v3.joblib |
XGBoost Ensemble | ~500 trees | 15 Feb 2026, 02:31 | Primary directional classifier |
nexus_lstm_v3.pth |
Bi-LSTM | ~2M | 14 Feb 2026, 11:45 | Sequence pattern recognition |
nexus_transformer_v2.pth |
Transformer (152M) | 5 epochs | 15 Feb 2026, 04:44 | Long-range dependency modeling |
nexus_medium_transformer_v1.pth |
Transformer (Medium) | 5 epochs | 15 Feb 2026, 05:49 | Balanced capacity/speed |
nexus_small_transformer_v1.pth |
Transformer (Small) | 10 epochs | 15 Feb 2026, 05:24 | Fast inference, high accuracy |
nexus_transformer_pretrained.pth |
Pretrained base | — | 14 Feb 2026, 07:22 | Foundation weights |
feature_scaler_v3.pkl |
StandardScaler | — | 15 Feb 2026, 02:31 | 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% |
🕐 Training Log
📈 Small Transformer — 10 epochs (15 Feb 2026)
| Epoch | Accuracy | Timestamp |
|---|---|---|
| 1 | 60.0% | 05:09 |
| 2 | 69.7% | 05:10 |
| 3 | 72.6% | 05:12 |
| 4 | 74.5% | 05:14 |
| 5 | 75.2% | 05:15 |
| 6 | 76.0% | 05:17 |
| 7 | 76.8% | 05:19 |
| 8 | 76.8% | 05:20 |
| 9 | 76.9% | 05:22 |
| 10 | 76.9% ✅ | 05:24 |
📈 Medium Transformer — 5 epochs (15 Feb 2026)
| Epoch | Accuracy | Timestamp |
|---|---|---|
| 1 | 58.1% | 05:34 |
| 2 | 69.8% | 05:37 |
| 3 | 72.7% | 05:41 |
| 4 | 74.8% | 05:45 |
| 5 | 76.2% ✅ | 05:49 |
📈 Nexus Transformer (152M) — 9 epochs (15 Feb 2026)
| Epoch | Accuracy | Timestamp |
|---|---|---|
| 1 | 51.3% | 06:30 |
| 2 | 52.4% | 06:51 |
| 3 | 52.4% | 07:12 |
| 4 | 53.1% | 07:32 |
| 5 | 54.6% | 07:52 |
| 6 | 55.3% | 08:13 |
| 7 | 57.3% | 08:33 |
| 8 | 58.1% | 08:54 |
| 9 | 58.7% ✅ | 09:14 |
Epoch 10 failed — weights from epoch 9 preserved.
⚡ Quick Start
Automatic (Recommended)
The Nexus Shadow-Quant app will auto-pull these models on first startup if no local models are found. Simply:
- Set your
HUGGINGFACE_TOKENandHF_REPO_IDin Settings. - Restart the backend.
- Models are downloaded and the predictor is ready instantly.
Manual
pip install huggingface_hub
huggingface-cli download Lukeed/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.
Dr. Nexus · Quantitative intelligence, engineered locally.