# Stock 15-Minute Return Prediction

Experiments for predicting stock 15-minute returns using high-frequency features.

## Data

- **Features**: alpha158 computed on 1-minute data
- **Target**: 15-minute forward returns (close[t+16]/close[t+1]-1)
- **Normalization**: industry, cs_zscore, or dual

## Notebooks

| Notebook | Purpose |
|----------|---------|
| `01_data_exploration.ipynb` | Load and explore 15m data structure |
| `02_baseline_model.ipynb` | Train baseline XGBoost model |

## Methodology

1. Load 1-minute kline data via Polars lazy frames
2. Compute/retrieve alpha158 features
3. Calculate 15-minute forward returns
4. Apply normalization (industry-neutralized or cross-sectional z-score)
5. Train gradient boosting models
6. Evaluate with IC and backtest