Python experiments that train neural networks on historical stock-price features and generate forecasts and diagnostic plots. The code is intended for learning and research; its output is not validated trading performance.
Use a Python environment compatible with the pinned NumPy, pandas, SciPy, and scikit-learn versions in requirements.txt (Python 3.11 is a practical starting point). TensorFlow availability depends on your platform.
git clone https://github.com/Neel-Sh/Stock-Prediction.git
cd Stock-Prediction
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txtpython train_model.py
python predict_realtime.pyBoth scripts prompt for a ticker, using AAPL when input is empty. Train first, then predict using the same ticker and the generated model artifacts. Training downloads market data and currently requests 100 epochs for the ensemble; it can take significant time and resources.
| File | Purpose |
|---|---|
train_model.py |
Downloads data, computes features, prepares sequences, trains and saves models |
predict_realtime.py |
Loads saved artifacts and produces forecasts and charts |
requirements.txt |
Python dependencies |
model_analysis.png |
Checked-in example visualization; not a current benchmark |
Training writes ticker-prefixed .h5 models and .pkl scaler, feature, and metadata files. These generated artifacts stay ignored. Only load pickle/model files you trust.
Read preprocessing and split logic before interpreting reported accuracy: scaling is fitted during feature preparation, before the later train/test split, so holdout metrics can include information leakage. Historical fit, heuristic confidence scores, and buy/sell labels do not establish out-of-sample returns. External data availability can also change independently of this project.
See Contributing for development checks. No repository-wide license file is included.