Third-Party Research & Methodology Only

This section shares summaries of third-party academic research and descriptions of quantitative models. The content represents the findings of the original researchers, not the opinions or recommendations of Foxholm Financial. Foxholm Financial does not publish hypothetical or backtested performance metrics on its quantitative research pages. All content is restricted to methodology, signal construction, factor logic, and risk architecture. SEC rules require that investment advisers not present misleading performance data, and our methodology-only approach reflects that standard and the firm's fiduciary obligations.

GARCH Volatility Model

Risk Model Time Series Volatility Forecasting

Robert Stowe, AAMS® Investment Advisor

GARCH (Generalized Autoregressive Conditional Heteroskedasticity) is a statistical model that captures how market volatility (the size of price swings) changes over time. Unlike simple volatility measures that treat every day the same, GARCH recognizes a fundamental pattern in financial markets: large price moves tend to be followed by more large moves, and calm periods tend to be followed by more calm periods.

This clustering behavior is one of the most well-documented features of financial return data. If the market drops 3% today, the chance of another large move (up or down) tomorrow is higher than usual. A model that assumes constant volatility misses this pattern entirely. GARCH solves this by updating its volatility estimate each day based on what just happened, producing a forecast that responds to current market conditions rather than relying solely on long-run averages.

Conceptual Framework

The name GARCH describes exactly what the model does. "Conditional" means the volatility estimate depends on (is conditioned on) recent information. "Heteroskedasticity" is the statistical term for volatility that changes over time, as opposed to a constant level. "Autoregressive" means the model uses its own past estimates as inputs, creating a feedback loop where yesterday's volatility estimate helps predict today's.

Robert Engle introduced the original ARCH model in 1982, showing that volatility could be modeled as a function of past squared returns. Tim Bollerslev generalized it in 1986 by adding past volatility estimates as additional inputs, creating GARCH. Engle received the 2003 Nobel Prize in Economics for this work, which transformed how financial risk is measured and managed.

Core Assumptions

GARCH models rest on assumptions about how volatility behaves. These assumptions are generally well-supported by data but have important limitations:

Volatility clusters: The foundational assumption is that volatility is persistent. Today's volatility level is informative about tomorrow's. This is strongly supported by empirical evidence across equity, fixed income, currency, and commodity markets.
Mean reversion: GARCH assumes volatility eventually returns to a long-run average level. After a spike (like a market crash), volatility gradually declines back toward normal. This assumption holds well over moderate time horizons but can break down during structural shifts in market conditions, such as a prolonged crisis where the "normal" level itself changes.
Parametric distribution: The basic GARCH model assumes returns follow a specific distribution (typically normal) after adjusting for time-varying volatility. In practice, returns have fatter tails than the normal distribution predicts, even after the GARCH adjustment. Extensions like GARCH with Student-t errors address this by using distributions that allow for more extreme events.
Symmetric response: The standard GARCH(1,1) model treats positive and negative shocks equally: a 2% gain and a 2% loss contribute the same amount to tomorrow's volatility forecast. In reality, negative shocks tend to increase volatility more than positive shocks of the same size, a phenomenon called the leverage effect. Asymmetric variants like EGARCH and GJR-GARCH address this.

Model Architecture

The GARCH model estimates volatility through a recursive process. Each day's volatility forecast depends on three components that are combined using estimated weights.

Input 1

Long-Run Average

Input 2

Yesterday's Shock

Input 3

Yesterday's Variance

Output

Today's Forecast

The GARCH(1,1) Equation

The standard GARCH(1,1) model expresses today's variance as a weighted combination of three terms:

\(\sigma^2_t = \omega + \alpha \cdot \epsilon^2_{t-1} + \beta \cdot \sigma^2_{t-1}\)

In plain language: today's predicted variance (σ²_t) equals a baseline constant (ω) plus a weight (α) times yesterday's squared return shock (ε²_t-1) plus another weight (β) times yesterday's variance estimate (σ²_t-1). The α parameter controls how quickly the model reacts to new shocks. The β parameter controls how much persistence (memory) the model retains from past estimates.

For financial data, typical parameter values are α around 0.05 to 0.10 and β around 0.85 to 0.95. The sum α + β measures overall persistence; values close to 1.0 indicate that volatility shocks decay slowly, which is the norm for equity markets. The long-run average variance equals ω / (1 - α - β).

Parameter Estimation

GARCH parameters are estimated using maximum likelihood estimation (MLE). The method finds the parameter values that make the observed return data most probable under the model. This requires specifying a distribution for the return innovations (the normal distribution is the default, but Student-t or generalized error distributions often fit better).

The estimation requires a reasonable amount of historical data. A minimum of two to three years of daily returns is typical, though more data generally produces more stable parameter estimates. The tradeoff is that very long estimation windows may include structural breaks (periods where market behavior fundamentally changed), which can distort the parameter estimates.

Asymmetric Variants

The standard GARCH model treats positive and negative return shocks identically. In practice, equity markets exhibit the leverage effect: negative returns increase volatility more than positive returns of the same magnitude. Several asymmetric extensions address this:

EGARCH (Exponential GARCH): Proposed by Nelson (1991), EGARCH models the log of variance, which naturally allows for asymmetric effects and ensures the variance is always positive without requiring parameter constraints. The model includes a term that differentiates the impact of positive and negative shocks.
GJR-GARCH: Proposed by Glosten, Jagannathan, and Runkle (1993), this variant adds a dummy variable that activates when the return is negative. The additional parameter directly measures the extra volatility impact of negative shocks compared to positive ones.
TGARCH (Threshold GARCH): Similar in concept to GJR-GARCH, the threshold model defines different volatility responses above and below zero return, capturing the asymmetry through separate coefficient regimes.

Risk Architecture

GARCH models are widely used as inputs to risk management systems. The quality of the volatility forecast directly affects the accuracy of risk measures like Value at Risk and the reliability of portfolio optimization outputs.

Model Risk

The primary risk is specification error: choosing the wrong GARCH variant or the wrong distribution for the innovations. A standard GARCH(1,1) with normal errors may significantly underestimate the probability of extreme events because it misses both the asymmetric volatility response and the fat tails of the return distribution. Using a GJR-GARCH with Student-t errors addresses both issues but adds parameters that must be estimated, which requires more data and introduces estimation uncertainty.

A second risk is structural breaks. GARCH models are estimated over a specific historical period, and the parameters reflect the average volatility dynamics during that period. If the market enters a fundamentally different regime (a new regulatory environment, a pandemic, a financial crisis without historical precedent), the estimated parameters may not apply. Rolling or expanding estimation windows partially address this but introduce their own tradeoffs in responsiveness versus stability.

Known Limitations

Limitations to Consider

Short-horizon focus: GARCH forecasts are most accurate one day ahead. Forecast accuracy degrades rapidly as the horizon extends. Multi-step-ahead forecasts converge toward the unconditional (long-run average) variance, meaning the model adds progressively less value over longer horizons.
Single-asset limitation: The basic GARCH model estimates volatility for one asset at a time. Portfolio risk requires modeling how multiple assets co-move, which demands multivariate extensions like DCC-GARCH (Dynamic Conditional Correlation). These models are substantially more complex and can be computationally intensive for large portfolios.
Jumpiness during regime transitions: When a quiet market suddenly becomes volatile (or vice versa), GARCH models react with a lag because they update incrementally. The model may underestimate risk at the start of a crisis and overestimate it after the crisis has passed, because the exponential smoothing takes time to adjust.
Parameter instability: GARCH parameters estimated over different sample periods can vary meaningfully. The sensitivity of the parameters to the estimation window means that two analysts using the same model but different data ranges may produce different volatility forecasts.
No structural interpretation: GARCH models describe volatility dynamics statistically but do not explain why volatility changes. They capture patterns in the data without providing insight into the economic mechanisms driving those patterns. This limits their usefulness for understanding whether current volatility conditions are likely to persist or reverse.

Practical Considerations

Risk Management Applications

GARCH volatility forecasts serve as inputs to several downstream risk applications:

Value at Risk (VaR): GARCH-based VaR replaces the assumption of constant volatility with a conditional estimate that reflects current market conditions. During high-volatility periods, the VaR estimate increases; during calm periods, it decreases. This produces more responsive risk limits than historical VaR with a fixed lookback window.
Option pricing: The Black-Scholes model assumes constant volatility, which is known to be incorrect. GARCH-based option pricing models allow volatility to evolve stochastically, producing pricing surfaces that better match observed market prices, particularly for out-of-the-money options where the volatility smile is most pronounced.
Portfolio optimization: Mean-variance optimization requires a covariance matrix as input. Using GARCH-estimated volatilities (and, with multivariate extensions, time-varying correlations) produces a covariance matrix that reflects current conditions rather than a fixed historical average. This can improve portfolio allocation during periods of elevated or depressed volatility.

Model Selection and Diagnostics

Choosing among GARCH variants requires balancing model complexity against fit. The standard diagnostic approach involves:

Information criteria: AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) penalize models for additional parameters while rewarding better fit. Lower values indicate a better balance of parsimony and explanatory power.
Standardized residual analysis: If the model is correctly specified, the standardized residuals (actual returns divided by the GARCH standard deviation) should behave like independent draws from the assumed distribution. Remaining autocorrelation or fat tails in the standardized residuals indicate misspecification.
Out-of-sample forecast evaluation: The ultimate test is whether the model produces useful forecasts on data it has not seen. Common evaluation metrics include mean squared error of variance forecasts and the Mincer-Zarnowitz regression of realized variance on forecasted variance.

Multivariate Extensions

Portfolio risk management requires modeling how assets co-move, not just individual volatilities. The DCC-GARCH model, introduced by Engle (2002), estimates time-varying correlations between assets while keeping the number of parameters manageable. It works in two stages: first, estimate individual GARCH models for each asset; second, model the time-varying correlation structure of the standardized residuals.

DCC-GARCH captures the well-documented tendency for correlations to increase during market stress. When stocks fall sharply, correlations between stocks tend to rise, reducing the diversification benefit of the portfolio precisely when it matters most. A correlation model that reflects this dynamic produces more realistic portfolio risk estimates than one that assumes fixed correlations.