From GARCH to Neural Network for Volatility Forecast

Pengfei Zhao, Haoren Zhu, Wilfred Siu Hung Ng, Dik Lun LeeAAAI 2024 · DOI

Introduction

Predicting financial volatility—the uncertainty in asset returns—is critical for risk management and option pricing. Historically, this task has been divided between two distinct fields: econometrics, which favors interpretable GARCH models, and machine learning, which utilizes powerful but often opaque neural networks. This paper bridges this divide by establishing a formal mathematical equivalence between these two paradigms, leading to a hybrid GARCH-LSTM model that leverages both domain expertise and computational flexibility.

Bridging the Gap: GARCH-NN Equivalence

The core contribution of this research is the proof that traditional GARCH models have direct neural counterparts. By constraining inputs and removing standard activation functions, the recursive structure of a GARCH(1,1) model becomes mathematically identical to a basic Recurrent Neural Network (RNN).

Structural equivalence between RNN/CNN architectures and GARCH variants.

As shown in the figure above, the authors establish several key relationships:

  • GARCH(1,1) ≈ RNN: Shares the same recursive hidden state logic as a basic RNN.
  • GJR-GARCH ≈ RNN with Sign Info: Captures market asymmetry (where bad news has a greater impact than good news) by processing the sign of previous returns.
  • FI-GARCH ≈ CNN: The long-term memory of Fractionally Integrated GARCH maps to a 1D Convolutional Neural Network where filter weights represent decaying historical impacts.

The authors propose a GARCH-LSTM architecture that "infuses" this GARCH logic directly into the LSTM framework. Furthermore, they replace standard Mean Squared Error (MSE) loss with Maximum Likelihood Estimation (MLE) based on Student’s t-distributions to better model the "fat tails" (extreme events) typical of financial data.

Experimental Evaluation

The models were tested on a decade of daily data (2012–2022) across five major asset classes: stock indexes (S&P 500, DJI, NASDAQ), exchange rates (EUR-USD), and commodities (Gold). Realized volatility, calculated via a 5-day rolling average, served as the ground truth for evaluation.

Key Results and Insights

The hybrid GARCH-LSTM consistently outperformed both pure econometric and pure deep learning models across various forecasting horizons.

  • Accuracy Gains: The model achieved a 3% improvement in Mean Absolute Error (MAE) and a 10% improvement in MSE over the next best performing models.
  • Reliability: The hybrid approach exhibited significantly lower variance (≈ 1e-4) compared to standalone neural networks (≈ 1e-3), making it more suitable for production trading environments.
  • Domain Knowledge Matters: Hard-coding "stylized facts" like volatility clustering into the network architecture proved more effective than relying on a deep model to learn these patterns from scratch. Interestingly, complex Transformer-based models often underperformed, highlighting the importance of temporal-aware architectures in finance.

Conclusion

By formalizing the relationship between GARCH and neural networks, this work provides a blueprint for integrating financial domain knowledge into deep learning. Future research may extend this "GARCH kernel" to Graph Neural Networks to model cross-asset correlations or adapt the framework to handle extreme market regime shifts.

Summary updated

BibTeX
@inproceedings{zhao2024from,
  author = {Pengfei Zhao and Haoren Zhu and Wilfred Siu Hung Ng and Dik Lun Lee},
  title = {{From GARCH to Neural Network for Volatility Forecast}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i15.29643},
}
Figure