Trading signal prediction pipeline that combines technical indicators, LLM-based sentiment analysis, and gradient-boosted classifiers to forecast short-term equity price direction.
The pipeline fetches daily OHLCV data, engineers 25+ technical features, generates a sentiment time-series (optionally scored by Claude via the Anthropic API), trains an XGBoost binary classifier with walk-forward cross-validation, and backtests a long/flat strategy on the held-out window.
OHLCV data ─── technical features ──┐
├── XGBoost ── backtest ── charts
headlines ─── LLM sentiment ────────┘
- No lookahead leakage. Train/test split is purely temporal; walk-forward CV mirrors how the model would be retrained in production.
- Signal lag. Positions are entered one day after signal generation to reflect real execution latency.
- Sentiment modelling. An Ornstein–Uhlenbeck mean-reverting process converts point-in-time headline scores into a realistic daily series, avoiding i.i.d. assumptions that would inflate the feature's apparent value.
git clone https://github.com/vagif693/quantmind.git
cd quantmind
pip install -e ".[dev]"# demo mode — no API key required
python -m quantmind.cli --ticker SPY --years 3
# with LLM sentiment scoring
export ANTHROPIC_API_KEY="sk-ant-..."
python -m quantmind.cli --ticker AAPL --years 5 --use-ai
# run tests
make test| Flag | Default | Description |
|---|---|---|
--ticker |
SPY |
Ticker symbol |
--years |
3 |
Years of data |
--horizon |
5 |
Forward-return window (days) |
--use-ai |
off | Enable Claude sentiment scoring |
--api-key |
env var | Anthropic API key |
--output-dir |
output/ |
Chart directory |
--no-charts |
off | Skip chart generation |
-v |
off | Debug logging |
quantmind/
├── src/quantmind/
│ ├── cli.py # entry point
│ ├── config.py # pipeline configuration (frozen dataclass)
│ ├── data.py # market data + synthetic fallback
│ ├── features.py # technical indicators, target construction
│ ├── sentiment.py # LLM scoring + OU process
│ ├── model.py # XGBoost training, walk-forward CV
│ ├── backtest.py # long/flat simulation, metrics
│ └── charts.py # matplotlib visualisations
├── tests/
│ └── test_pipeline.py
├── output/ # generated charts
├── pyproject.toml
└── Makefile
Sample output on 3 years of SPY data (demo mode):
| Metric | Value |
|---|---|
| Sharpe | −0.60 |
| Max drawdown | −9.3% |
| Win rate | 50.8% |
| Alpha vs B&H | −3.0% |
The negative alpha is expected in demo mode — synthetic sentiment carries
limited signal. With --use-ai and real headline data the sentiment features
contribute meaningfully (see feature importance chart). Regardless, the
purpose of this project is to demonstrate pipeline architecture, not to present
a production trading system.
- Data: yfinance, pandas, numpy
- Indicators: ta
- Sentiment: Anthropic Claude API
- ML: XGBoost, scikit-learn
- Visualisation: matplotlib, seaborn
- Testing: pytest
- Packaging: pyproject.toml, setuptools
This is an educational project. It is not financial advice and should not be used for real trading without substantial additional validation.
MIT