Modeling Stochastic Volatility with the Kalman Filter

Volatility is a cornerstone of quantitative finance, essential for risk management, option pricing, and portfolio construction. While models like GARCH and its variants have been the workhorses for volatility modeling, they have limitations. Stochastic volatility (SV) models offer a more flexible and realistic alternative, and the Kalman filter provides a effective framework for their estimation.

Limitations of GARCH Models

Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models treat volatility as a deterministic function of past returns and past volatilities. This implies that, given the historical data, the current volatility is known with certainty. This is a restrictive assumption, as it does not account for the inherent randomness in volatility itself. Furthermore, GARCH models can be slow to react to sudden changes in market conditions.

Stochastic Volatility Models

Stochastic volatility models address these limitations by treating volatility as an unobserved, latent variable that follows its own stochastic process. A common formulation for a simple SV model is:

• Return Equation: r_t = σ_t ε_t
• Volatility Equation: log(σ_t²) = α + β log(σ_t-1²) + η_t

where:

• r_t is the asset return at time t.
• σ_t is the stochastic volatility.
• ε_t and η_t are independent, zero-mean Gaussian noise terms.

The logarithm of the squared volatility is modeled as an autoregressive process, which means that the volatility itself is a random variable.

State-Space Representation and the Extended Kalman Filter

To apply the Kalman filter, we need to cast the SV model into a state-space form. The state equation is straightforward from the volatility equation:

• State Equation: h_t = α + β h_t-1 + η_t, where h_t = log(σ_t²) is the hidden state.

The measurement equation, however, is non-linear:

• Measurement Equation: y_t = log(r_t²) = log(σ_t² ε_t²) = h_t + log(ε_t²)

The term log(ε_t²) is the measurement noise. Since the standard Kalman filter is designed for linear systems, we need to use a non-linear extension, the Extended Kalman Filter (EKF).

The EKF linearizes the non-linear measurement equation at each time step using a first-order Taylor series expansion. The algorithm then proceeds in the same predict-update cycle as the standard Kalman filter, but using the linearized model.

The Unscented Kalman Filter for Improved Accuracy

While the EKF is a common choice, it can introduce significant errors if the system is highly non-linear. The Unscented Kalman Filter (UKF) offers a more accurate alternative. Instead of linearizing the model, the UKF uses a deterministic sampling technique called the unscented transform to capture the mean and covariance of the state distribution. A set of "sigma points" are chosen, propagated through the non-linear function, and then used to reconstruct the transformed distribution. The UKF generally provides better estimates than the EKF for non-linear systems, at a comparable computational cost.

Practical Applications

By applying the Kalman filter (or its non-linear variants) to a stochastic volatility model, we can obtain a real-time, filtered estimate of the latent volatility. This has several practical applications:

• Improved Risk Management: A more accurate and responsive measure of volatility leads to better Value-at-Risk (VaR) and other risk calculations.
• Enhanced Option Pricing: The Black-Scholes model assumes constant volatility. By feeding a real-time volatility estimate from a Kalman filter into the Black-Scholes formula, we can obtain more accurate option prices.
• Volatility-Based Trading Strategies: Trading strategies can be designed to capitalize on changes in volatility. For example, a strategy might go long volatility when the Kalman filter indicates it is low and expected to rise.

In conclusion, the Kalman filter provides a robust and flexible framework for estimating stochastic volatility models. By treating volatility as a latent, random variable, these models offer a more realistic representation of market dynamics than traditional GARCH models. The resulting real-time volatility estimates can be a valuable input for a wide range of trading and risk management applications.