Articles
GARCH Interview Questions: Volatility Clustering, Forecasts, and Diagnostics
Practice GARCH interview questions with a worked variance recursion, shock forecasts, return-unit checks, residual diagnostics, and horizon alignment.
A return series has almost no visible autocorrelation, but large moves keep arriving near other large moves. An interviewer asks why a constant-variance model misses this pattern. GARCH answers a question about conditional variation, not necessarily about the direction of the next return.
A useful approach to GARCH interview questions is to label the forecast quantity, calculate the variance recursion, and diagnose what remains in standardized residuals. A fitted curve that follows past turbulence is not sufficient evidence of useful future forecasts.
Evidence boundary: Official arch and statsmodels documentation establishes the model and API facts below. The parameter example, horizon ledger, and diagnostic decisions are original exercises and inference. No candidate reports, hiring-frequency claims, or investment recommendations are made. Begin with Derive Annualized Portfolio Volatility to distinguish variance from volatility and its units.
What does GARCH model?
Official model fact: A standard constant-mean GARCH(1,1) writes a return as a mean plus an innovation, with a variance recursion involving the previous squared innovation and previous conditional variance. The arch documentation describes this formulation in its forecasting guide.
Use explicit notation:
r_t = mu_t + epsilon_t
epsilon_t = sqrt(h_t) * z_t
h_t = omega + alpha * epsilon_(t-1)^2 + beta * h_(t-1)
For the basic formulation here, standardized innovations have conditional mean zero and conditional variance one. The mean model determines mu_t; the volatility model determines h_t. The square root of h_t is conditional standard deviation.
The recursion uses an innovation, not automatically the raw return. If the mean is nonzero or has dynamics, subtract the fitted conditional mean first. Inserting raw squared returns changes the model.
Original interpretation: Alpha controls the immediate response to the previous squared innovation. Beta carries forward the previous conditional variance. A negative return and an equally large positive return have the same squared innovation in this symmetric formulation. Modeling different responses to their signs requires a different specification.
Can you calculate a shock and its forecast path?
Original exercise: Use variance units of squared percentage points. Set omega to 0.1, alpha to 0.1, and beta to 0.8. At forecast origin t, the current conditional variance is one and the observed innovation is three percentage points.
h_(t+1) = 0.1 + 0.1 * 3^2 + 0.8 * 1 = 1.8
This is the one-step conditional variance forecast. Its conditional standard deviation is approximately 1.342 percentage points. The squared shock contribution is 0.9; treating the innovation as a signed value would give a different and incorrect recursion.
For later horizons, future innovations are not observed. Under the model's conditional second-moment assumptions, substitute the forecast variance for the expected squared future innovation. Do not repeatedly insert the already observed shock of three into every future period.
| Horizon from the same origin | Forecast variance | Conditional standard deviation |
|---|---|---|
| 1 | 1.800 | 1.342 |
| 2 | 1.720 | 1.311 |
| 3 | 1.648 | 1.284 |
| 4 | 1.5832 | 1.258 |
Model derivation: Here the persistence coefficient alpha plus beta is 0.9. The finite unconditional variance implied by this stationary second-moment calculation is 0.1 / (1 - 0.9) = 1. Thus the horizon-h forecast follows:
E_t[h_(t+h)] = 1 + 0.9^(h-1) * (1.8 - 1)
This equation describes mean reversion of a variance forecast from one fixed information set. It does not predict that each realized future variance or return will follow this smooth path.
What conditions are needed for the long-run variance?
For the usual nonnegative GARCH(1,1) parameterization, omega is positive and alpha and beta are nonnegative. With standardized innovations of variance one, a stationary solution with a finite unconditional second moment requires alpha plus beta below one. State these assumptions before presenting the familiar long-run variance formula.
If alpha plus beta equals one, that finite-variance formula divides by zero. If it exceeds one, the same algebra does not justify a positive stationary long-run variance. A numerical optimizer returning parameters is not proof that the moment assumptions you need are satisfied.
Official specification fact: arch provides GARCH-family processes, including asymmetric terms and different powers. Their parameterization must be read before transferring the simple squared GARCH formula. See the GARCH process API.
Second-moment stationarity and strict stationarity are different concepts. Do not claim that alpha plus beta below one exhausts every possible stationarity question. For this interview exercise, the narrower claim concerns the finite variance used in the expectation calculation.
A parameter estimate very near the boundary also deserves uncertainty analysis. Small changes can materially change the implied long-run variance and persistence. Report estimation context and uncertainty rather than treating an extreme point estimate as a physical constant.
Why do return units matter so much?
Original unit exercise: A one-percent return can be represented as 0.01 in decimal units or 1 in percentage-point units. Multiplying the return and innovation by 100 multiplies variance by 10,000 and standard deviation by 100.
For the same model expressed consistently, alpha and beta are dimensionless while omega and the variance state change with squared units. Comparing two output numbers without checking their input units can create a false disagreement of four orders of magnitude.
The example's forecast variance of 1.8 squared percentage points corresponds to 0.00018 in squared decimal-return units. Its standard deviation is approximately 0.01342 in decimal units, matching approximately 1.342 percentage points.
Engineering inference: Record frequency, return definition, missing dates, corporate-action handling where relevant, and units before fitting. A daily close-to-close return series and intraday observations need different interpretation. Counting observations is not automatically the same as counting calendar days.
Do not silently fill missing returns with zero. That inserts artificial calm observations and changes both the mean and volatility evidence. Decide how to handle missing records from the actual data-generating and collection process.
How do you diagnose a fitted model?
Original diagnostic matrix: After fitting the mean and variance models, calculate standardized residuals as the fitted innovation divided by fitted conditional standard deviation. Examine both their values and their squares.
| Observed residual pattern | Possible explanation | Next check |
|---|---|---|
| Standardized residuals retain serial correlation | Mean dynamics or other dependence remain | Revisit the mean specification and data timing |
| Squared standardized residuals retain dependence | Conditional variance structure remains | Check volatility specification and lags |
| Extreme tails remain frequent | Innovation distribution or tail fit is inadequate | Inspect tail diagnostics and held-out coverage |
| Positive and negative shocks behave differently | Symmetric variance response is inadequate | Compare an appropriate asymmetric specification |
| One period drives the fit | Outlier, regime shift, or data error | Inspect source records and stability across windows |
This is a diagnostic decision aid, not a one-to-one proof of causation. A rejected test may have several explanations. A nonrejected test does not establish that all dependence has disappeared or that every forecast horizon is calibrated.
Official diagnostic API fact: statsmodels exposes a Ljung-Box test for autocorrelation and accepts lag and model-degree-of-freedom choices. See acorr_ljungbox. State what series you test, which lags you choose, and how fitted parameters affect the reference calculation.
Checking raw squared returns alone answers a different question from checking squared standardized residuals. The latter asks whether the fitted conditional variance absorbed the relevant structure. Always label the input to a diagnostic plot or test.
Heavy tails and volatility clustering are also different properties. An independent heavy-tailed sequence can have occasional extreme values without persistent variance dynamics. A volatility model with normal standardized innovations can generate changing conditional variance while imposing its own conditional tail assumptions.
How should forecasts be aligned and evaluated?
Official API fact: arch distinguishes forecast mean, variance, and residual variance outputs. Forecast alignment can refer to the origin or the target. For mean models with dynamics, variance and residual variance can differ at longer horizons. See the forecast-output documentation.
Engineering inference: Save both the information cutoff and the target period. A row labeled with the forecast origin is not necessarily a prediction for that same row's realized return. Misaligned evaluation can compare a forecast with data it already used.
Use rolling or expanding historical training windows according to the intended application. At each origin, use only information available then. Reestimating parameters with the entire sample before describing forecasts as historical out-of-sample evidence introduces future information.
Conditional variance is latent, so a realized squared innovation is a noisy comparison target. Choose an evaluation measure and benchmark deliberately. Compare with a constant-variance or other justified baseline, and inspect whether improvement survives across relevant time windows.
For simulation forecasts, state the innovation distribution, number of paths, random seed, and Monte Carlo uncertainty. Simulation error and parameter uncertainty are separate. Increasing path count does not remove a misspecified model or uncertain fitted coefficients.
Is a multi-day forecast just daily volatility times a square root?
Original reasoning: Under a constant-mean model with conditionally uncorrelated future innovations, the variance of their sum can be obtained by summing the relevant expected future innovation variances. The volatility of that aggregate is the square root of the sum.
In the original ledger, the first three variance forecasts sum to 5.168. The standard deviation of the three-period innovation sum under those stated assumptions is approximately 2.273 percentage points. Multiplying the first period's volatility by the square root of three gives approximately 2.324, a different number because the forecast variances change across horizons.
If the mean model introduces additional dynamics, or the requested object differs from an innovation sum, revisit the covariance and forecast-error calculation. Do not transfer a square-root-of-time shortcut without stating its assumptions.
Likewise, E[sqrt(h)] is not generally sqrt(E[h]). A reported square root of a variance forecast is a particular summary, not automatically the expected future realized volatility. Name the quantity you calculate.
What does successful optimization establish?
Engineering inference: Inspect the optimizer status, parameter constraints, likelihood value, and initialization sensitivity. A successful termination message describes the numerical search; it does not establish that the chosen distribution or volatility specification fits the application.
Compare plausible initializations when the fit behaves unexpectedly, and inspect whether missing observations or an extreme return dominate the result. Preserve the data revision and fitted configuration so the comparison can be reproduced. Changing the mean specification while leaving the residual interpretation unchanged can invalidate the diagnostic story.
If parameters lie near a boundary, explain which uncertainty calculation was used and whether its approximations are credible there. An interview answer can acknowledge this limitation without claiming that every difficult fit requires abandoning GARCH. Identify the specific assumption or numerical behavior that needs checking.
Five questions to practice variance reasoning
These verified destinations cover volatility, features, simulation, and evaluation. They are related practice questions; no claim is made that their descriptions report a direct GARCH interview round.
| Question | GARCH follow-up |
|---|---|
| Derive Annualized Portfolio Volatility | Track variance, standard deviation and frequency. |
| Design Features for Residual Volatility | Define residuals before modeling their variance. |
| Estimate Volatility from Gaussian and Brownian Samples | Separate estimator uncertainty from variance dynamics. |
| Design and backtest a trading strategy | Prevent future information in forecast evaluation. |
| Implement simulation-based portfolio optimizer in Python | Separate path simulation from parameter estimation. |
Continue with Design and backtest a trading strategy. Explain the information cutoff and benchmark before interpreting a fitted volatility model's historical results.