GARCH & Volatility Modeling Services
Financial volatility is not constant—it clusters, spikes, and reacts asymmetrically to bad news. The GARCH family is built to capture exactly this, modelling and forecasting time-varying variance. MAS Research specifies, estimates, and validates volatility models across the full family, from single-asset GARCH to multivariate DCC and realized-volatility approaches.
GARCH (generalized autoregressive conditional heteroskedasticity) models the time-varying volatility of financial returns, capturing volatility clustering—the tendency of large changes to be followed by large changes. Extensions handle asymmetry (EGARCH, GJR-GARCH), multiple assets (DCC-GARCH), and mixed-frequency drivers (GARCH-MIDAS), while realized-volatility methods use high-frequency data. They are central to risk, pricing, and volatility forecasting.
What GARCH and volatility models do
Financial returns have a distinctive signature: periods of calm are punctuated by bursts of turbulence, so large changes tend to be followed by large changes and small by small. This volatility clustering means the variance of returns is not constant over time—it is conditional on recent history. Ordinary models that assume constant variance miss this entirely, mis-stating risk exactly when it matters most.
The GARCH family (generalized autoregressive conditional heteroskedasticity) is built precisely to model this time-varying variance. A GARCH model lets today’s volatility depend on recent shocks and recent volatility, capturing clustering directly and producing volatility forecasts that respond to market conditions. From this foundation, a family of extensions handles the other stylised facts of returns: asymmetry—the leverage effect, where bad news raises volatility more than equally sized good news—through EGARCH, TGARCH, and GJR-GARCH; co-movement across assets through multivariate DCC-GARCH; and long-run volatility driven by macro or mixed-frequency variables through GARCH-MIDAS. Where high-frequency data are available, realized-volatility measures offer a complementary, data-rich route.
When to use which model
Use a standard GARCH when the goal is to model and forecast the volatility of a single return series and clustering is the main feature. Move to an asymmetric model (EGARCH, GJR-GARCH, TGARCH) when negative and positive shocks plausibly affect volatility differently—almost always the case for equities, where the leverage effect is strong. Use DCC-GARCH when the question involves several assets and how their volatilities and correlations move together over time (essential for portfolio risk and correlation dynamics). Use GARCH-MIDAS when you want to link long-run volatility to lower-frequency drivers such as macroeconomic conditions, and realized volatility when high-frequency intraday data let you measure volatility more directly.
The choice is not cosmetic: fitting a symmetric model to data with a strong leverage effect, or a univariate model to an inherently multivariate risk question, gives the wrong answer. We select the member of the family that matches the stylised facts of your data and the question—and test that the chosen model actually fits.
The GARCH family
| Model | Captures | Use when |
|---|---|---|
| ARCH / GARCH | Volatility clustering | Single-series volatility, the baseline |
| EGARCH / GJR / TGARCH | Asymmetry (leverage effect) | Bad news raises volatility more (e.g. equities) |
| DCC-GARCH | Time-varying correlations across assets | Multi-asset risk & correlation dynamics |
| GARCH-MIDAS | Long-run volatility from low-frequency drivers | Linking volatility to macro conditions |
| Realized volatility | Volatility measured from high-frequency data | Intraday data available |
Specifying and validating a volatility model
Sound volatility modelling follows a sequence. First, the mean equation is specified and the returns checked for ARCH effects—there is no point fitting GARCH if the data show no conditional heteroskedasticity, so this is tested, not assumed. The error distribution is chosen to match the fat tails of financial returns: a normal distribution usually understates tail risk, so Student-t or skewed-t distributions are typically more appropriate. The model order (the GARCH lag structure) is selected with information criteria and kept parsimonious—a simple specification often forecasts better than an elaborate one.
Validation is what separates a credible volatility model from a fitted curve. Standardised residuals are checked for remaining ARCH effects and autocorrelation—if the model is adequate, these should be gone. For forecasting and risk use, the model is evaluated out of sample, and where it feeds a risk measure such as Value-at-Risk, backtesting confirms whether realised exceptions match the model’s predictions. For multivariate models, the estimated correlation dynamics are checked for plausibility and the model’s conditions respected. We report the diagnostics and the distributional and specification choices, because a volatility figure is only as trustworthy as the model that produced it.
Match the model to the stylised facts—and validate it. Equity returns show a leverage effect, so a symmetric GARCH understates risk after bad news; financial returns have fat tails, so a normal distribution understates extremes. We test for ARCH effects, choose a fat-tailed distribution, check standardised residuals, and backtest risk use.
Software
We deliver the GARCH family in established, reproducible tools—R (rugarch, rmgarch) and Python (arch), plus EViews/Stata—covering univariate and multivariate (DCC) models, asymmetric and MIDAS variants, realized-volatility measures, fat-tailed distributions, out-of-sample evaluation, and Value-at-Risk backtesting, all with versioned code.
How we deliver a volatility-modeling study
Volatility modelling sits within our wider financial-econometrics practice—so the model matches the data’s stylised facts, the distribution fits the tails, and the diagnostics are reported in full.
We start from your returns data and the question—single-asset or multi-asset, in-sample description or out-of-sample forecasting, standalone or feeding a risk measure. We specify the mean equation, test for ARCH effects, select the family member (symmetric, asymmetric, multivariate, MIDAS) and a fat-tailed distribution, and estimate the model, then validate it through residual diagnostics and—where relevant—out-of-sample and backtesting evaluation.
Reporting sets out the specification, the distributional choice, the family member and why it was chosen, the estimates (including asymmetry and, for multivariate models, correlation dynamics), and the diagnostics—so the volatility estimates and forecasts can be judged.
You receive the estimated volatility model with its parameters and interpretation, conditional-volatility and (where relevant) correlation series, out-of-sample forecasts and backtesting results where the use case requires them, the full diagnostics, and reproducible analytical code and analysis-ready files (where appropriate and permitted).
Volatility modeling across finance research
Time-varying volatility is central to asset pricing, risk, and markets—so the GARCH family runs throughout finance and financial-economics research.
Asset Pricing & Markets
Modelling and forecasting return volatility, the volatility–return relationship, and risk premia across equities, bonds, and other assets.
Risk Management
Volatility as the engine of Value-at-Risk and other risk measures, with backtesting to confirm the model holds up.
Portfolio & Correlation Dynamics
Multivariate DCC-GARCH for time-varying correlations—central to diversification, hedging, and portfolio risk.
Energy & Commodity Markets
Volatility and its drivers in oil, gas, and commodity returns, often with asymmetric and MIDAS specifications.
Cryptocurrency & Digital Assets
Modelling the extreme, clustered volatility of crypto returns with appropriate fat-tailed, asymmetric models.
Macro-Finance
Linking long-run volatility to macroeconomic conditions via GARCH-MIDAS, connecting markets to the economy.
GARCH & volatility: common questions
Modeling or forecasting financial volatility?
Whether it is single-asset volatility, multi-asset correlation dynamics, or volatility feeding a risk measure, we fit the right member of the GARCH family—with the correct distribution, asymmetry where it matters, and the diagnostics and backtesting that make the result trustworthy.