VolLab

Inspiration

Financial markets move through periods of calm and chaos, but detecting regime shifts in volatility is notoriously difficult. Traditional statistical methods struggle with nonlinear patterns and sudden jumps. We wanted to explore whether modern machine learning can improve regime detection and help traders, funds, and regulators anticipate market risks before they escalate.

What it does

VolLab is a machine learning system for volatility regime detection. We first generate synthetic market data using stochastic models (Geometric Brownian Motion, Merton jump-diffusion, and Heston stochastic volatility) to capture different market behaviors. Then we train supervised ML models — LightGBM and TabNet — to classify volatility regimes. Finally, we combine them with both blending and stacking ensembles, achieving robust predictions across different simulated environments. Brought together in a aesthetic dashboard via streamlit.

How we built it

Data Simulation: Built a synthetic dataset with GBM, Merton, and Heston processes to capture distributional and jump-risk dynamics. Feature Engineering: Extracted rolling-window features covering distributional stats, technical indicators, dependence metrics, and risk-based measures. ML Modeling: Trained LightGBM (gradient boosting trees) and TabNet (deep learning for tabular data). Ensembling: Implemented both blending and stacking ensembles for stronger generalization. Visualization: Plotted volatility regime shifts to demonstrate interpretable model outputs.

Challenges we ran into

Balancing realism and tractability when simulating financial data. Feature engineering for volatility is non-trivial — many features correlated too strongly or failed under regime shifts. Blending vs. stacking required experimentation with meta-learners to avoid overfitting.

Accomplishments that we're proud of

Successfully implemented three stochastic models to simulate realistic financial regimes. Built a hybrid ensemble system that outperformed individual models in classification accuracy. Created interpretable visualizations of regime changes — turning abstract financial math into something you can “see.”

What we learned

How to combine financial theory with machine learning for real-world problems. The trade-offs between boosting methods and neural networks for tabular finance data. That ensembling is extremely powerful when models capture different inductive biases.

What's next for VolLab

Integrating real market data (e.g., SPY, VIX, crypto) for live testing. Exploring reinforcement learning to adapt trading strategies to detected regimes.

Built With

• python
• sklearn
• streamlit