Stock Price Prediction with LSTM

Stock Price Prediction with LSTM is a hands-on deep learning project that demonstrates how sequential models can be applied to real-world financial data. Using historical OHLCV (Open, High, Low, Close, Volume) data, the project builds and trains an LSTM network to capture time-dependent patterns in stock movements.

The pipeline handles everything from preprocessing and sliding-window dataset creation to model training with early stopping and evaluation. The results are presented with intuitive visualizations — training and validation loss curves, predicted vs. actual stock prices, and short-horizon forecasts into the future. Metrics such as RMSE, MAE, and MAPE provide quantitative insight into performance.

This project serves as both a learning tool and a portfolio-ready showcase of time-series forecasting, deep learning, and financial modeling with PyTorch.

---

Features

• Load stock data from CSV or fetch with Yahoo Finance (via yfinance)
• Preprocessing: scaling & sliding window dataset creation
• LSTM model with dropout and Adam optimizer
• Metrics: RMSE, MAE, MAPE
• Plots:
  • Training & validation curves
  • Predicted vs actual prices
  • Short-horizon future forecast
• Saved artifacts: best_lstm.pt, scaler.pkl, metrics.json

---

Project Structure

stock-lstm-forecasting/
├─ README.md
├─ LICENSE
├─ requirements.txt
├─ data/
│  ├─ fetch_yfinance.py      # Fetch data from Yahoo Finance
│  └─ aapl.csv               # Stock dataset (real or synthetic)
├─ src/
│  ├─ train_lstm_stock.py    # Training script
│  ├─ evaluate.py            # Evaluation script
│  └─ utils.py               # Helpers (scaling, metrics, windowing)
└─ outputs/
   ├─ best_lstm.pt
   ├─ scaler.pkl
   ├─ metrics.json
   ├─ training_curves.png
   ├─ predicted_vs_actual.png
   └─ future_forecast.png

---

Setup

python -m venv .venv
# Windows
.venv\Scripts\activate
# Linux/macOS
source .venv/bin/activate

pip install -r requirements.txt

---

Fetch Data (optional)

# downloads daily OHLCV for AAPL (Jan 2015 → today)
python data/fetch_yfinance.py --ticker AAPL --start 2015-01-01 --out data/aapl.csv

Or use the included synthetic dataset (data/aapl.csv).

---

Train the Model

python src/train_lstm_stock.py --input data/aapl.csv --column close     --lookback 60 --epochs 25 --batch-size 64 --outdir outputs --horizon 1 --seed 42

---

Evaluate the Model

python src/evaluate.py --input data/aapl.csv --model outputs/best_lstm.pt     --column close --lookback 60 --horizon 1 --outdir outputs

---

Results

Training & Validation Loss

---

Predicted vs Actual

---

Short-Horizon Forecast

---

Metrics (metrics.json):

{
  "rmse": 3.59,
  "mae": 3.59,
  "mape": 1.97
}

---

Recommendations

• Train longer (50–100 epochs) for improved stability
• Try multi-step forecasts (--horizon 5 or --horizon 30)
• Experiment with other assets (e.g., MSFT, GOOGL, TSLA)
• Add more features (Volume, technical indicators)