LSTM Volatility Prediction — URA Uranium ETF

A deep learning approach to forecasting 21-day forward realised volatility on the URA (Global X Uranium ETF) using stacked LSTM recurrent neural networks — benchmarked against a classical GARCH(1,1) baseline.

---

Overview

Volatility is one of the hardest signals to forecast in financial markets — it clusters, jumps, and reacts non-linearly to regime shifts. This project tackles that challenge on the URA uranium ETF (an asset with a particularly volatile history) by training a stacked LSTM to learn temporal patterns in price-based features and predict realised volatility 21 trading days ahead.

The model is benchmarked head-to-head against GARCH(1,1), the classical econometric workhorse, on the same out-of-sample window.

---

Highlights

• End-to-end Jupyter notebook — from raw OHLCV download to model evaluation and report-ready plots.
• Stacked LSTM (64 → 32 units) with dropout regularisation and a custom loss function combining MSE with a variance-matching penalty.
• 8 engineered features: log returns, rolling means/stds (5- and 21-day), volume change, intraday range, and open-close range.
• GARCH(1,1) baseline computed via the arch library for a fair classical-vs-deep-learning comparison.
• Adaptive training — EarlyStopping + ReduceLROnPlateau prevent overfitting on limited financial data.
• Reproducible — fixed random seeds and chronological train/val/test split (70/15/15).

---

Results

Out-of-sample test set (Nov 2023 → Mar 2026, 516 sequences):

Model	R²	RMSE	MAE
LSTM (ours)	0.033	0.00649	0.00506
GARCH(1,1)	−0.158	0.00710	—

The LSTM outperforms GARCH(1,1) by ~0.19 in R² on the test set, demonstrating that nonlinear temporal patterns provide meaningful signal beyond classical econometric models.

A full discussion — including residual analysis, training dynamics, and limitations — is in URA_Report_Final.pdf.

---

Tech Stack

Layer	Tools
Language	Python 3.10+
Deep learning	TensorFlow / Keras
Classical model	arch (GARCH)
Data	yfinance, pandas, numpy
ML utilities	scikit-learn
Visualisation	matplotlib
Environment	Jupyter Notebook

---

Architecture

Input (60 timesteps × 8 features)
   │
   ├─ LSTM(64, return_sequences=True)
   ├─ Dropout(0.2)
   ├─ LSTM(32)
   ├─ Dropout(0.2)
   ├─ Dense(16, relu)
   └─ Dense(1, linear) ──► 21-day forward realised volatility

Total parameters: 31,649
Loss: MSE + 0.3 × variance-matching penalty
Optimizer: Adam (lr = 1e-4)

---

Getting Started

1. Clone

git clone https://github.com/GeorgeLazos/ura_volatility_prediction_LTSM.git
cd ura_volatility_prediction_LTSM

2. Install dependencies

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

pip install numpy pandas matplotlib scikit-learn tensorflow yfinance arch jupyter

3. Run the notebook

jupyter notebook ura_volatility_prediction_LSTM.ipynb

The notebook downloads URA data live from Yahoo Finance, so an internet connection is required on the first run.

---

Project Structure

.
├── ura_volatility_prediction_LSTM.ipynb   # Main notebook (data → model → eval)
├── URA_Report_Final.pdf                   # Full written report
├── LICENSE                                # MIT
├── .gitignore
└── README.md

---

Methodology

1. Data acquisition — Daily OHLCV for URA from inception (2010-11-05), via yfinance.
2. Feature engineering — Log returns, rolling means/stds, volume change, intraday/open-close ranges.
3. Target — 21-day forward realised volatility (rolling std of log returns).
4. Sequence construction — 60-day sliding windows.
5. Scaling — StandardScaler fit on training set only (no leakage).
6. Training — Adam, MSE + variance-penalty loss, EarlyStopping (patience 20), ReduceLROnPlateau (factor 0.5, patience 5).
7. Evaluation — RMSE, MAE, R² on out-of-sample test set; head-to-head vs GARCH(1,1).

---

License

Released under the MIT License — Copyright © 2026 Georgios Lazos.

---

Author

Georgios Lazos
MSc AI & Machine Learning — Deep Learning Final Project
GitHub @GeorgeLazos