Self-hosted LLM-powered crypto sentiment & price prediction terminal
Track 5 coins in real-time with Reddit/news sentiment (local Mistral 7B via llama-cpp-python), on-chain whale intelligence, XGBoost/Prophet/LSTM ensemble predictions, and composite trading signals — all on your own hardware. No cloud LLM fees.
Reddit/News → Sentiment (Mistral 7B GGUF + FinBERT) ↘
Binance OHLCV → Technical Indicators (40+ features) → Ensemble Signals → Streamlit Dashboard
Etherscan On-Chain → Whale Flow Analysis ↗
- Docker Desktop installed
- 8GB+ RAM (16GB recommended for Mistral 7B)
- GPU optional but recommended (CUDA 12 supported)
- Python 3.10+ (if running without Docker)
git clone https://github.com/AaravMehta-07/CryptoNerve
cd CryptoNerve
cp .env.example .env
# Edit .env with your API keys (Reddit, NewsAPI, Etherscan)The GGUF model (~3.5 GB) is not included in this repo (excluded via .gitignore).
Place it at models/mistral-7b-instruct-v0.1.Q3_K_M.gguf or set the LLM_MODEL_PATH env variable.
You can download it from TheBloke on HuggingFace.
docker compose up -dhttp://localhost:8501
pip install -r requirements.txt
python -m pipeline # or python orchestrator.py
streamlit run dashboard/app.pyNote:
llama-cpp-pythonrequires a C++ compiler. For GPU support:
CMAKE_ARGS="-DLLAMA_CUDA=on" pip install llama-cpp-python==0.2.90
| Service | Free Tier | Use |
|---|---|---|
| Yes | Social sentiment | |
| NewsAPI | 100 req/day | News sentiment |
| Etherscan | 5 req/sec | Whale tracking |
All keys are optional — the system degrades gracefully with warnings if any are missing.
| Page | Description |
|---|---|
| 📊 Live Dashboard | Multi-coin price cards, Fear & Greed, real-time signals |
| 📈 Price & Technicals | OHLCV + RSI + MACD + Bollinger Bands |
| 💬 Sentiment Analysis | Sentiment timeline, FUD/bullish composition, narrative detection |
| 🐋 On-Chain Intelligence | Whale flows, exchange inflow/outflow, accumulation detection |
| 🤖 AI Predictions | Ensemble prediction history with accuracy tracking |
| ⚡ Signals & Alerts | Full signal breakdown with component scores |
| ⏳ Backtesting | Strategy backtest with equity curve, Sharpe ratio, drawdown |
| 📝 AI Reports | LLM-generated market reports (Mistral 7B) |
| 🔍 Explainability | SHAP waterfall + feature importance visualization |
| 📊 Model Performance | Accuracy tracking across models, coins, horizons |
| Interval | Task |
|---|---|
| 1 min | Price data (Binance) |
| 2 min | On-chain whale data (Etherscan) |
| 5 min | Reddit posts (PRAW) + Signal generation |
| 10 min | News articles (NewsAPI / CryptoCompare) + Sentiment analysis |
| 6 hours | Model retraining |
| Model | Weight | Role |
|---|---|---|
| XGBoost | 45% | Primary classifier with 40+ features incl. sentiment + technical + on-chain |
| Prophet | 30% | Time-series trend, seasonality, holiday effects |
| LSTM | 25% | Sequential patterns via 48-step sliding window |
Additional ML: LightGBM, AutoGluon (tabular), Optuna hyperparameter optimization, SHAP explainability.
NLP: Mistral 7B (GGUF, CPU/GPU via llama-cpp-python) + FinBERT for financial sentiment.
# ML
xgboost, lightgbm, tensorflow, scikit-learn, autogluon, optuna, shap
# NLP / LLM
transformers, torch, llama-cpp-python (GGUF)
# Data / On-chain
praw, newsapi-python, web3, aiohttp
# Dashboard
streamlit, plotly, wordcloud
# DB
psycopg2, sqlalchemy (PostgreSQL with connection pooling)
This is for educational purposes only. Not financial advice.
Crypto markets are highly volatile. Never invest more than you can afford to lose.
Crypto Sentinel v1.1 — Built with ❤️ using Python, Streamlit, PostgreSQL, and llama-cpp-python