Volatility Clustering and GARCH-X Neural Hybrids

Volatility exhibits clustering: large price moves tend to be followed by more large moves, not necessarily in the same direction. This clustering is captured by GARCH models, which have been standard in finance for decades. However, GARCH has limitations: it assumes symmetric response to shocks, and may miss complex nonlinearities. Hybrid approaches combining GARCH structure with neural networks capture both classical dynamics and complex patterns.

GARCH: The Classical Approach

GARCH models volatility as a function of recent squared returns (shocks) and recent volatility:

σ²(t) = ω + α × r²(t-1) + β × σ²(t-1)

This captures volatility clustering: if recent returns were large (high r²), future volatility is elevated. If volatility was recently high, it tends to stay high.

GARCH is interpretable, tractable, and works reasonably well. However:

  • It assumes linear relationships (response to shocks is proportional)
  • It treats all large shocks identically (whether up or down)
  • It misses regime-switching dynamics
  • It ignores relevant external information (news, volatility in other assets)

GARCH-X: Incorporating Exogenous Variables

Natural extension: include exogenous variables (realized volatility, option-implied volatility, macroeconomic surprises):

σ²(t) = f(r²(t-1), σ²(t-1), exogenous variables)

This is GARCH-X. By including more information, forecasts improve. However, choosing relevant exogenous variables and their functional form requires domain expertise.

Neural Network Enhancements

Neural networks can learn the functional form automatically. A hybrid approach:

σ²(t) = α × ReLU(DNN([r(t-1), σ(t-1), exogenous])) + β × σ²(t-1)

The neural network component learns complex relationships between inputs and volatility, while the β × σ²(t-1) term ensures strong mean-reversion (volatility does not explode).

Alternatively, a more flexible hybrid:

σ²(t) = GARCH([r(t-1), σ(t-1)]) + DNN_residual([exogenous]) + ε

The GARCH component captures classical dynamics; the neural network component captures residual patterns not explained by GARCH.

LSTM-Based Approaches

RNNs naturally capture temporal dependencies. An LSTM volatility model:

  • Takes as input a sequence of recent returns, volatility levels, and exogenous variables
  • Maintains hidden state encoding the current volatility regime
  • Outputs predicted volatility for the next period

LSTMs can capture volatility clustering automatically without explicit GARCH terms.

Regime-Switching Hybrids

Volatility exhibits regime-switching: normal periods with moderate volatility vs crisis periods with high volatility. Hybrid models can include regime-switching:

  • Low-volatility regime: GARCH with lower persistence
  • High-volatility regime: GARCH with higher persistence
  • Neural network determines which regime is active based on current data

Multi-Asset Volatility Models

Volatility in one asset depends on volatility in related assets. Extending the model to multiple assets simultaneously:

  • Capture within-asset dynamics via GARCH-type terms
  • Capture cross-asset spillovers via neural network that learns interdependencies

This is especially important for options portfolios with exposure to multiple underlyings.

Training and Validation

Volatility models are typically evaluated via likelihood (does the model assign high likelihood to realized volatility in test data?) or forecast error (how accurate are multi-step predictions?).

Walk-forward validation is crucial: train on period 1-100, test on 101-110, then retrain on 1-110 and test on 111-120, etc. This prevents overfitting.

Practical Use Cases

  • Risk management: volatility forecasts drive position limits and margin requirements
  • Options pricing: local volatility models and stochastic volatility models use these forecasts
  • Volatility trading: predict volatility changes to profit from variance swaps or options strategies
  • Hedging: adapt hedging ratios based on predicted volatility evolution

Conclusion

Combining classical GARCH structure with neural network flexibility creates powerful volatility models that capture both known stylized facts and complex patterns. These hybrids often outperform pure GARCH or pure neural approaches, demonstrating the value of domain-informed machine learning in derivatives.