9.5 Volatility modeling

Concept of volatility

Volatility measures how much a financial asset's price moves around over time. It's central to financial mathematics because nearly every pricing model, risk metric, and trading strategy depends on getting volatility right. This topic covers how to measure it, model it, and use it in practice.

Definition and importance

Volatility quantifies the dispersion of returns for a security or market index. A stock that swings 5% daily is far more volatile than one that moves 0.5%, and that difference matters for everything from option pricing to portfolio construction.

• Higher volatility means greater price fluctuations and more investment risk
• Option pricing models require a volatility input to produce fair values
• Portfolio managers use volatility estimates to set position sizes and allocate across assets
• Regulators and risk teams rely on volatility to determine capital requirements

Historical vs. implied volatility

These are two fundamentally different ways to think about volatility, and you need to understand both.

Historical volatility is backward-looking. You take a window of past returns (say, 30 or 60 trading days), compute their standard deviation, and annualize it. It tells you what volatility was.

Implied volatility is forward-looking. You take an observed option price, plug it into a pricing model like Black-Scholes, and solve for the volatility that makes the model price match the market price. It tells you what the market expects volatility to be.

Traders compare the two constantly. If implied volatility is much higher than recent historical volatility, options may be "expensive." If it's lower, they may be "cheap." Neither measure is objectively better; they answer different questions.

Volatility clustering

Volatility clustering is the empirical observation that large price changes tend to be followed by large price changes, and small changes tend to follow small changes. You can see this clearly on any stock chart: calm periods and turbulent periods tend to persist.

• This pattern shows up across asset classes and time frames
• It directly contradicts the assumption of constant volatility used in basic models like Black-Scholes
• Volatility clustering is the primary motivation behind ARCH and GARCH models, which explicitly allow volatility to change over time
• Risk managers need to account for clustering because a single large move often signals more large moves ahead

Statistical measures of volatility

These foundational measures give you the quantitative tools to assess and compare volatility across assets. They're also the building blocks for every more complex model covered later.

Standard deviation

Standard deviation measures how spread out a set of returns is from its mean. In finance, it's the most common single-number summary of an asset's risk.

$\sigma = \sqrt{\frac{\sum_{i=1}^{n} (x_i - \mu)^2}{n - 1}}$

where $\sigma$ is the standard deviation, $x_i$ are individual return observations, $\mu$ is the mean return, and $n$ is the number of observations. The $n - 1$ denominator (Bessel's correction) is used for sample data.

A stock with an annualized standard deviation of 30% is roughly twice as volatile as one at 15%. To annualize a daily standard deviation, multiply by $\sqrt{252}$ (since there are about 252 trading days per year).

Variance

Variance is simply the square of standard deviation:

$\sigma^2 = \frac{\sum_{i=1}^{n} (x_i - \mu)^2}{n - 1}$

Why care about variance separately? Because many models (GARCH, variance swaps, portfolio optimization) work directly with variance rather than standard deviation. Variance is also additive for uncorrelated assets, which makes portfolio math cleaner.

The tradeoff is interpretability: variance is in "squared return" units, which aren't intuitive. That's why standard deviation is preferred for reporting.

Coefficient of variation

The coefficient of variation (CV) lets you compare volatility across assets with very different average returns:

$CV = \frac{\sigma}{\mu} \times 100\%$

For example, if Asset A has a mean return of 10% with a standard deviation of 15%, its CV is 150%. If Asset B has a mean return of 2% with a standard deviation of 4%, its CV is 200%. Even though Asset B has lower absolute volatility, it's riskier relative to its return. This makes CV useful when comparing across asset classes like equities vs. bonds.

Time series models

The statistical measures above treat volatility as a fixed number. Time series models recognize that volatility changes over time and try to capture those dynamics.

ARCH models

The Autoregressive Conditional Heteroskedasticity (ARCH) model, introduced by Robert Engle in 1982, was the first formal framework for time-varying volatility. The core idea: today's volatility depends on how large recent price shocks were.

An ARCH(q) model defines conditional variance as:

$\sigma_t^2 = \alpha_0 + \alpha_1\epsilon_{t-1}^2 + \alpha_2\epsilon_{t-2}^2 + \cdots + \alpha_q\epsilon_{t-q}^2$

where $\sigma_t^2$ is the conditional variance at time $t$, $\epsilon_{t-i}^2$ are past squared residuals (shocks), and $\alpha_0 > 0$, $\alpha_i \geq 0$ are parameters.

After a large shock (positive or negative), $\epsilon_{t-1}^2$ is large, so $\sigma_t^2$ increases. This is exactly how the model captures volatility clustering. The limitation is that you often need many lagged terms (large $q$) to fit the data well, which leads to many parameters.

GARCH models

The Generalized ARCH (GARCH) model, developed by Bollerslev in 1986, solves the parameter problem by adding lagged variance terms. A GARCH(p,q) model is:

$\sigma_t^2 = \alpha_0 + \sum_{i=1}^q \alpha_i\epsilon_{t-i}^2 + \sum_{j=1}^p \beta_j\sigma_{t-j}^2$

The $\beta_j\sigma_{t-j}^2$ terms allow past variance to feed into current variance, giving the model "memory." In practice, GARCH(1,1) fits most financial return series remarkably well:

$\sigma_t^2 = \alpha_0 + \alpha_1\epsilon_{t-1}^2 + \beta_1\sigma_{t-1}^2$

The sum $\alpha_1 + \beta_1$ controls persistence. When it's close to 1, volatility shocks decay slowly, which matches what we see in real markets.

EGARCH and GJR-GARCH

Standard GARCH treats positive and negative shocks symmetrically: a +3% return and a -3% return produce the same effect on future volatility. In reality, negative shocks tend to increase volatility more than positive shocks of the same size. This is called the leverage effect.

EGARCH (Nelson, 1991) models this asymmetry using a logarithmic specification:

$\ln(\sigma_t^2) = \omega + \sum_{i=1}^q (\alpha_i |z_{t-i}| + \gamma_i z_{t-i}) + \sum_{j=1}^p \beta_j \ln(\sigma_{t-j}^2)$

where $z_t$ are standardized residuals. The $\gamma_i$ parameters capture asymmetry: if $\gamma_i < 0$, negative shocks increase volatility more than positive ones. A key advantage is that the log form guarantees $\sigma_t^2 > 0$ without requiring parameter constraints.

GJR-GARCH (Glosten, Jagannathan, and Runkle, 1993) takes a different approach, adding an indicator variable that "switches on" extra volatility when returns are negative. Both models are widely used in equity markets where the leverage effect is pronounced.

Stochastic volatility models

GARCH-family models treat volatility as a deterministic function of past data. Stochastic volatility models go further by treating volatility itself as a random process with its own source of uncertainty. This added flexibility is especially valuable for option pricing.

Heston model

The Heston model (1993) is the most widely used stochastic volatility model for equity options. It specifies two coupled stochastic differential equations:

Asset price: $dS_t = \mu S_t \, dt + \sqrt{v_t} \, S_t \, dW_t^S$

Variance process: $dv_t = \kappa(\theta - v_t) \, dt + \sigma \sqrt{v_t} \, dW_t^v$

where:

• $S_t$ is the asset price, $v_t$ is the instantaneous variance
• $\kappa$ is the speed of mean reversion, $\theta$ is the long-run variance level
• $\sigma$ is the volatility of volatility ("vol of vol")
• $W_t^S$ and $W_t^v$ are Wiener processes with correlation $\rho$

The mean-reverting variance process pulls $v_t$ toward $\theta$ over time, which matches the empirical observation that volatility doesn't drift off to infinity or zero. The correlation $\rho$ (typically negative for equities) captures the leverage effect. A major practical advantage is that the Heston model has a semi-closed-form solution for European option prices, making calibration feasible.

SABR model

The Stochastic Alpha Beta Rho (SABR) model is the standard for interest rate derivatives, particularly swaptions and caps/floors.

Forward rate: $dF_t = \alpha_t F_t^\beta \, dW_t^1$

Volatility: $d\alpha_t = \nu \alpha_t \, dW_t^2$

where $F_t$ is the forward rate, $\alpha_t$ is the stochastic volatility, $\beta \in [0,1]$ controls the backbone (how volatility relates to the level of rates), $\nu$ is the vol of vol, and the two Brownian motions have correlation $\rho$.

The SABR model is popular because Hagan et al. (2002) derived an accurate closed-form approximation for implied volatility as a function of strike, which makes it fast to calibrate and easy to interpolate across strikes. The $\beta$ parameter is particularly useful: $\beta = 1$ gives lognormal dynamics, $\beta = 0$ gives normal dynamics, and values in between blend the two.

Hull-White model

The Hull-White model is primarily an interest rate model rather than a pure volatility model, but it's included here because it addresses how short-rate volatility is modeled in fixed income.

$dr_t = [\theta(t) - a \, r_t] \, dt + \sigma \, dW_t$

where $r_t$ is the short rate, $\theta(t)$ is a time-dependent drift function, $a$ is the mean reversion speed, and $\sigma$ is the volatility.

The time-dependent $\theta(t)$ is calibrated to exactly fit the current term structure of interest rates, which is a significant advantage over earlier models like Vasicek. This makes it practical for pricing interest rate derivatives where matching today's yield curve is essential.

Volatility surfaces

A volatility surface is a three-dimensional plot of implied volatility against strike price and time to expiration. It captures information that a single volatility number cannot, and it's essential for pricing and hedging options accurately.

Term structure of volatility

The term structure describes how implied volatility varies across different expiration dates for options at the same moneyness (e.g., at-the-money).

• An upward-sloping term structure means longer-dated options have higher implied volatility, suggesting the market expects uncertainty to increase over time
• A downward-sloping term structure means near-term options are more expensive in volatility terms, often seen when a specific event (earnings, election) is approaching
• A humped term structure can occur when a near-term event creates elevated short-term vol that's expected to normalize

The term structure matters for calendar spreads and for anyone comparing options across different maturities.

Volatility smile

The volatility smile is the pattern of implied volatility across strike prices for options with the same expiration. If you plot implied vol against strike, you often see a U-shaped curve: both deep in-the-money and deep out-of-the-money options have higher implied volatilities than at-the-money options.

This directly contradicts Black-Scholes, which assumes a single constant volatility for all strikes. The smile exists because:

• Markets price in the possibility of large jumps (fat tails) that Black-Scholes ignores
• Supply and demand for specific strikes shifts implied vol
• The smile shape varies by asset class: currency options tend to show a more symmetric smile, while equity options show a pronounced skew

Volatility skew

The volatility skew is an asymmetric version of the smile, most prominent in equity markets. Out-of-the-money puts typically have significantly higher implied volatility than out-of-the-money calls at the same distance from the current price.

This negative skew reflects:

• Demand for downside protection (portfolio insurance via puts)
• The empirical observation that markets crash more often than they spike
• The leverage effect: falling prices increase a firm's effective leverage, raising volatility

In commodity markets, you sometimes see a positive skew, where upside calls are more expensive. This can happen when supply disruptions create fear of price spikes. The steepness of the skew is itself a tradeable quantity and contains information about market sentiment.

Volatility indices

Volatility indices distill the information in options markets into a single number, making it easy to track market-implied volatility over time.

VIX index

The VIX, published by the Chicago Board Options Exchange (Cboe), measures the 30-day expected volatility of the S&P 500 implied by option prices. It's often called the "fear gauge" because it tends to spike during market selloffs.

• The VIX is calculated from a wide strip of S&P 500 option prices across many strikes, not just at-the-money options
• It's quoted in annualized percentage points: a VIX of 20 means the market expects roughly 20% annualized volatility over the next 30 days
• Typical range: 12-20 in calm markets, 30+ during stress, and it briefly exceeded 80 during the 2020 COVID crash
• VIX futures and options are actively traded, enabling direct volatility exposure

Other volatility indices

Several analogous indices exist for other markets:

• VSTOXX: implied volatility of EURO STOXX 50 options (European equities)
• VKOSPI: expected volatility in the Korean stock market
• GVZ: 30-day expected volatility of gold prices
• MOVE: implied volatility of U.S. Treasury options (fixed income equivalent of VIX)

These indices allow cross-market volatility comparisons and serve as underlyings for volatility derivatives in their respective markets.

Volatility trading strategies

These strategies aim to profit from changes in volatility itself, rather than from directional moves in the underlying asset.

Options strategies

Options are the primary tool for expressing volatility views because their prices are directly tied to implied volatility.

• Long straddle: buy a call and a put at the same strike and expiration. You profit if the underlying moves significantly in either direction. This is a bet that realized volatility will exceed implied volatility.
• Long strangle: similar to a straddle but with different strikes (out-of-the-money call and put). Cheaper to enter but requires a larger move to profit.
• Iron condor: sell an out-of-the-money call spread and an out-of-the-money put spread simultaneously. You profit if the underlying stays within a range, which is a bet that realized volatility will be lower than implied.
• Butterfly spread: profits from the underlying staying near a specific price. Like the condor, it benefits from low realized volatility.

Delta-neutral strategies hedge out directional exposure so that the position's P&L depends primarily on volatility.

Volatility swaps

A volatility swap is an OTC contract where the payoff depends directly on realized volatility versus a pre-agreed strike:

$\text{Payoff} = N \times (\sigma_{\text{realized}} - \sigma_{\text{strike}})$

where $N$ is the notional amount. If realized volatility over the contract period comes in at 25% and the strike was 20%, the long side receives $N \times 5\%$.

Volatility swaps provide pure volatility exposure with no directional risk and no need to delta-hedge. The challenge is that they're harder to replicate and hedge than variance swaps, which is why variance swaps are more liquid.

Variance swaps

Variance swaps work the same way but are based on variance (squared volatility):

$\text{Payoff} = N \times (\sigma_{\text{realized}}^2 - \sigma_{\text{strike}}^2)$

Variance swaps are more commonly traded because they can be replicated with a static portfolio of options across all strikes, making them easier to price and hedge. The payoff is convex in volatility: large moves in the underlying have a disproportionately large impact on the payoff. This means a long variance swap position benefits more from a volatility spike than a long volatility swap would.

Volatility risk management

Managing volatility risk means identifying, measuring, and controlling your portfolio's sensitivity to changes in volatility.

Value at Risk (VaR)

Value at Risk estimates the maximum loss a portfolio is likely to experience over a given time horizon at a specified confidence level.

For example, a 1-day 99% VaR of $1 million means: "There is a 1% probability that the portfolio will lose more than $1 million in a single day."

Three common calculation methods:

1. Historical simulation: use actual past returns to build a loss distribution
2. Variance-covariance (parametric): assume returns are normally distributed and use the portfolio's standard deviation
3. Monte Carlo simulation: generate thousands of random scenarios based on assumed return distributions

VaR is widely used and required by regulators under Basel III, but it has real limitations. It says nothing about how bad losses could be beyond the threshold, and it can underestimate risk during regime changes because it typically relies on recent history.

Expected shortfall

Expected shortfall (also called Conditional VaR or CVaR) addresses VaR's biggest weakness by asking: "Given that we've exceeded the VaR threshold, what's the average loss?"

For example, a 95% expected shortfall of $2 million means the average loss in the worst 5% of scenarios is $2 million. This is always a larger number than the corresponding VaR and gives a better picture of tail risk.

Expected shortfall is a coherent risk measure (it satisfies mathematical properties like subadditivity that VaR does not), and the Basel Committee has moved toward requiring it for market risk capital calculations.

Stress testing

Stress testing simulates extreme scenarios to see how a portfolio would perform under conditions that standard models might not capture.

• Historical scenarios: replay past crises (2008 financial crisis, 2020 COVID crash, 1998 LTCM collapse) through current positions
• Hypothetical scenarios: construct plausible but extreme events like a sudden 50% volatility spike, a correlation breakdown, or a liquidity freeze
• Results help risk managers set appropriate limits, adjust hedges, and allocate capital for worst-case outcomes

Stress testing complements VaR and expected shortfall because those measures are based on "normal" market conditions. Stress tests specifically target the abnormal conditions where the biggest losses occur.

Applications in finance

Volatility modeling feeds into nearly every area of quantitative finance. Here are the three most direct applications.

Option pricing

Volatility is the single most important input in option pricing that isn't directly observable. In the Black-Scholes model, every other input (stock price, strike, rate, time) is known; only volatility must be estimated or implied.

• Higher implied volatility increases both call and put prices because greater uncertainty makes the option's potential payoff larger
• The volatility smile and skew mean that a single flat volatility assumption won't correctly price options across different strikes
• Practitioners use the full volatility surface to price exotic options and structured products, interpolating and extrapolating implied vol as needed

Portfolio management

Volatility estimates directly affect how portfolios are constructed and rebalanced.

• Risk parity strategies weight assets so that each contributes equally to total portfolio volatility, rather than allocating equal dollar amounts
• Volatility targeting dynamically adjusts leverage or exposure to maintain a constant level of portfolio volatility (e.g., scaling down equity exposure when the VIX rises)
• Mean-variance optimization requires a covariance matrix, which depends entirely on volatility and correlation estimates

Risk assessment

Volatility is embedded in the most common performance and risk metrics:

• Sharpe ratio = excess return / standard deviation. A higher Sharpe ratio means better risk-adjusted performance.
• Sortino ratio uses downside deviation instead of total standard deviation, focusing only on harmful volatility
• Volatility forecasts feed into risk budgeting, position limits, and capital adequacy calculations required by regulators

Limitations and challenges

Volatility models are powerful but imperfect. Understanding their weaknesses is just as important as understanding how they work.

Model risk

Different models can produce meaningfully different volatility estimates from the same data. A GARCH(1,1) might give you an annualized volatility of 18% while a stochastic volatility model gives 21%. Trading decisions based on one model can look very different from those based on another.

• No single model captures all features of real volatility dynamics
• Overreliance on any one model creates blind spots
• Best practice is to use multiple models, compare their outputs, and apply expert judgment where they disagree

Estimation issues

Volatility is a latent variable: you never observe it directly. You can only estimate it from price data, and those estimates are sensitive to your choices.

• The length of the estimation window matters: too short and estimates are noisy, too long and they're stale
• Outliers (a single extreme return) can dramatically shift estimates
• High-frequency data introduces microstructure noise (bid-ask bounce, discrete tick sizes) that inflates measured volatility
• Parameters estimated in one market regime may not hold in the next

Regime changes

Financial markets periodically shift between distinct volatility regimes: calm periods with low, stable volatility and crisis periods with high, erratic volatility. The transition between regimes is often abrupt and hard to predict.

• Standard GARCH models adapt gradually, so they're slow to recognize a regime shift and slow to revert after one ends
• Regime-switching models (like Markov-switching GARCH) explicitly model transitions between states but add complexity and require identifying the number of regimes
• The 2008 financial crisis and 2020 COVID crash are textbook examples: volatility jumped from historically low levels to extreme levels within days
• Building flexibility into your modeling approach, rather than relying on a single fixed model, is the most practical defense against regime changes