Heston Model

The Heston model is an option-pricing model in which an asset’s variance changes randomly over time and tends to revert toward a long-run level. It models the asset price and variance together and allows their shocks to be correlated, helping the model represent implied-volatility skew and term structure that a single constant-volatility Black-Scholes input cannot reproduce.

The model was introduced by Steven L. Heston in 1993. It is a valuation framework, not a prediction that prices or volatility will follow the assumed process exactly.

Key Takeaways

• The Heston model replaces constant volatility with a stochastic variance process.
• Variance, not volatility itself, follows the model’s square-root mean-reverting process.
• Correlation between asset-price and variance shocks helps produce asymmetric return distributions and implied-volatility skew.
• European option values can be obtained through the model’s characteristic function and numerical integration, often called a semi-closed-form approach.
• Pricing calibration usually fits risk-neutral parameters to current option prices or implied volatilities across strikes and maturities.
• A good in-sample fit does not prove that parameters are stable, that hedges will work, or that the model is reliable outside the calibrated region.
• The long-run variance level is (\theta); its square root is a volatility scale, not the expected future volatility itself.
• Mean reversion has a time scale: under constant parameters, the variance gap has model half-life (\ln(2)/\kappa).
• Risk-neutral parameters calibrated from option prices need not equal physical-measure parameters estimated from historical returns.
• Simulation schemes must handle the square-root variance process carefully because a naive Euler step can generate negative discrete values.
• Calibration should be assessed by residual pattern, parameter stability, bid-ask fit, and out-of-sample behavior rather than one optimizer score.

Risk-Neutral Heston Equations

Under a common risk-neutral specification, the asset price (S_t) and instantaneous variance (v_t) follow:

with correlated Brownian shocks:

Here, (r) is the risk-free rate under the simplified specification and (q) is a continuous dividend yield or carry term. Production models may use curves, discrete dividends, funding adjustments, or additional factors rather than constant (r) and (q).

What the Parameters Mean

Parameter effects interact. For example, a more negative (\rho) can strengthen downside skew under common equity calibrations, but the observed surface also depends on (v_0), (\kappa), (\theta), (\xi), maturity, and the forward inputs.

Expected Variance and Mean-Reversion Speed

Under the constant-parameter risk-neutral variance process, conditional expected variance is:

If (v_0) is above (\theta), expected variance declines toward the long-run level; if it is below (\theta), expected variance rises toward it. The model half-life of the variance gap is:

For example, (\kappa=2) per year implies a half-life of about (0.347) years, or roughly 4.2 months under a simple year conversion. This does not forecast when observed market volatility will halve. It describes the conditional expectation under one calibrated process.

The quantity (\sqrt{\theta}) is often described informally as long-run volatility. More precisely, (\theta) is long-run variance under the selected measure and parameterization. In general, the expected square root of variance is not equal to the square root of expected variance.

Why Stochastic Variance Matters

The basic Black-Scholes model applies one volatility assumption to an option under its simplified framework. In actual markets, options with different strikes and maturities commonly have different implied volatilities.

Heston addresses part of that mismatch by allowing variance to evolve randomly. Correlation links variance shocks to asset-price shocks. When price declines and variance increases tend to occur together, a negative correlation parameter can help represent the higher implied volatility often observed for lower-strike equity options.

This flexibility does not guarantee an exact fit. A one-factor diffusion model may still struggle with very short-dated skew, jumps, multiple volatility factors, or unusual surface shapes.

Risk-Neutral vs. Physical Parameters

The equations used to price options are commonly written under a risk-neutral measure (\mathbb{Q}). The asset drift is then linked to rates and carry, while option prices embed compensation for volatility risk through the risk-neutral variance dynamics.

A historical model under the physical measure (\mathbb{P}) addresses a different question: how the analyst estimates the real-world evolution of returns and variance. Its mean-reversion speed, long-run variance, and other parameters can differ from their risk-neutral counterparts.

Substituting historical parameters into a pricing formula without a market-price-of-risk framework can produce values inconsistent with current option prices. Conversely, interpreting a risk-neutral calibration as a literal return forecast confuses valuation with statistical prediction.

Heston vs. Black-Scholes

Black-Scholes remains useful for quoting implied volatility and comparing options in common units. Heston adds structure to the volatility surface; it does not make benchmark quoting obsolete.

Practical Example: Fitting a Volatility Skew

Assume three six-month put options on the same stock have these illustrative market-implied volatilities:

A Black-Scholes calculation using one constant volatility cannot match all three market prices at once. Using 24% for every strike understates the volatility embedded in the lower-strike put and overstates the volatility embedded in the higher-strike put.

A Heston calibration instead searches for (v_0), (\kappa), (\theta), (\xi), and (\rho) that reduce pricing or implied-volatility errors across the selected options. A negative (\rho), together with the other parameters, can help create the downward-sloping skew in this example.

The output should be judged by more than one total error number. An analyst should inspect errors by strike and maturity, compare prices rather than only percentages, test parameter stability, and check whether the fitted model produces reasonable prices outside the calibration set.

A five-parameter model can match many surface points without making each parameter precisely identified. For example, changes in (\kappa), (\theta), and (\xi) can partly offset one another over the available maturities. Stable prices can therefore coexist with unstable fitted parameters.

How Calibration Works

A pricing calibration commonly follows these steps:

1. Select liquid option quotes and align spot, forward, discount, dividend, timestamp, and settlement inputs.
2. Convert bid, ask, or midpoint evidence into the chosen calibration objective.
3. Choose parameter bounds, initial guesses, weights, and numerical tolerances.
4. Reprice the selected options under trial parameter sets.
5. Minimize weighted price or implied-volatility errors.
6. Review residuals, parameter plausibility, numerical stability, and out-of-sample behavior.

A price-error objective can be written schematically as:

where (\mathbf{p}=(v_0,\kappa,\theta,\xi,\rho)) and (w_i) is the selected weight for option (i). An implied-volatility objective replaces price differences with volatility differences.

The objective can place greater weight on liquid quotes, price errors, relative price errors, vega-weighted errors, or selected strikes and maturities. Different objectives can produce different parameters even from the same market snapshot.

No objective is universally best. The choice should match the intended use and be documented before interpreting the resulting parameters.

Estimating a physical-measure process from historical returns is a different task. Option pricing uses risk-neutral dynamics and embeds market prices of risk. Historical estimates should not be substituted mechanically for pricing parameters.

Calibration Diagnostics

A calibration review should examine:

• price and implied-volatility residuals by strike, moneyness, and maturity;
• model values relative to bid and ask, not only midpoint;
• repeated runs from different starting values;
• sensitivity to parameter bounds and quote filters;
• parameter changes between adjacent valuation dates;
• prices and Greeks for instruments excluded from the fit;
• static-arbitrage behavior of model-generated option prices;
• numerical integration or solver convergence; and
• reproducibility from archived inputs and code version.

A visually smooth fitted surface can still be unsuitable for a particular exotic payoff or hedge. Validation should connect diagnostics to the model’s intended use.

The Feller Condition

A commonly reviewed parameter condition is:

Under the continuous-time square-root process, this condition is sufficient to keep the variance process away from the zero boundary. Calibrations can violate it, and the mathematical process can still remain nonnegative, but boundary behavior and numerical implementation then require particular care.

The Feller condition is a diagnostic, not proof that a calibration is economically sound. A parameter set can satisfy the condition and still fit market prices poorly or produce unstable hedges.

Pricing and Numerical Methods

Heston’s original European-option solution uses a characteristic function and numerical integration. It is often described as closed form or semi-closed form because the expectation is reduced to one-dimensional numerical integrals rather than simulated path by path.

Other methods are used when the payoff or exercise feature requires them:

• Fourier inversion or fast Fourier transform variants for grids of European options;
• finite-difference methods for barriers, boundaries, and early-exercise features;
• trees or lattices under selected discretizations;
• Monte Carlo simulation for path-dependent or multi-factor structures; and
• approximations or hybrid models when speed is more important than full revaluation.

Implementation choices matter. Characteristic-function branch handling, integration limits, quadrature, discretization bias, preservation of nonnegative variance, and convergence tolerances can all affect results.

Two implementations using the same parameter set can disagree if these controls differ. Regression tests should include prices, Greeks, limiting cases, and difficult parameter regimes rather than only typical at-the-money options.

Simulation and the Variance Boundary

The continuous-time square-root process is nonnegative, but a naive Euler discretization can produce a negative variance step. Simply taking an absolute value, reflecting, or setting a negative value to zero changes the simulated process and can introduce bias.

Common implementation approaches include:

• full-truncation Euler schemes that modify how negative discrete values enter drift and diffusion;
• quadratic-exponential schemes designed around the conditional moments of variance;
• exact or near-exact sampling methods with greater computational complexity; and
• sufficiently refined time grids with convergence tests for the payoff being valued.

There is no single best scheme for every payoff and parameter region. Barrier monitoring, strong negative correlation, high vol-of-vol, Feller-condition violations, and long simulation horizons can expose different errors. Results should be benchmarked against a trusted European-option implementation where possible.

Uses in Valuation and Risk

• European option valuation: fit a stochastic-volatility model across strikes and maturities.
• Volatility-surface analysis: separate level, skew, curvature, and maturity effects through a parameterized process.
• Exotic option valuation: provide an underlying stochastic-volatility process for numerical pricing.
• Scenario analysis: revalue positions after changes in spot, variance, correlation, or term structure.
• Risk sensitivities: calculate model-based Greeks and volatility-factor exposures.

Using Heston for a product does not establish that the model is approved, calibrated, or adequate. Model governance should define intended use, limitations, market-data hierarchy, independent validation, reserves, and monitoring.

Greeks and Hedging Under Heston

Heston Greeks depend on both spot and the stochastic-variance state. A position can have:

• delta and gamma to the underlying;
• sensitivity to initial variance (v_0);
• exposure to mean reversion, long-run variance, vol-of-vol, and correlation;
• strike- and maturity-specific surface risk; and
• numerical sensitivity to the calibration and pricing method.

Recalibrating the model after a market move can produce a different risk result from holding parameters fixed. A risk report should state whether it presents frozen-parameter Greeks, recalibrated scenarios, or market-quote bucket sensitivities.

Delta hedging alone does not hedge variance risk. An option used to offset vega can add skew, maturity, gamma, and liquidity basis. A model fit that prices vanillas closely can still produce unstable exotic-option hedges because different stochastic-volatility dynamics can fit similar current prices.

Model Validation and Governance

1. Define approved products, payoffs, markets, and valuation or risk uses.
2. Verify equations, measure, parameter units, curves, dividends, and contract conventions.
3. Benchmark European prices against an independent characteristic-function implementation.
4. Test integration, finite-difference, and simulation convergence across difficult parameter sets.
5. Compare analytical and bump-and-revalue Greeks under documented bump sizes.
6. Challenge calibration weights, bounds, starting values, filters, and optimizer settings.
7. Review residuals by strike and maturity and compare values with bid-ask ranges.
8. Test out-of-sample prices, hedge behavior, and parameter stability.
9. Stress jumps, surface shifts, illiquid wings, boundary behavior, and model extensions.
10. Document limitations, reserves, overrides, monitoring thresholds, and independent review.

Risks and Limitations

• One-factor variance risk: one variance process may not capture short- and long-horizon volatility dynamics simultaneously.
• No-jump limitation: the basic diffusion model does not directly represent discontinuous price or variance jumps.
• Calibration instability: different starting values, bounds, quote sets, or weights can produce different parameters.
• Parameter-identification risk: several parameter combinations can generate similar option prices.
• Extrapolation risk: a good fit to liquid strikes and maturities may not extend to wings or long dates.
• Numerical risk: integration, discretization, simulation, and code errors can create inaccurate values or Greeks.
• Market-data risk: stale, crossed, illiquid, or asynchronous quotes can distort calibration.
• Hedge risk: model sensitivities can change rapidly and realized markets can move outside assumed dynamics.
• Product mismatch: early exercise, discrete dividends, barriers, funding, or multiple underlyings may require extensions.
• Measure confusion: risk-neutral parameters can be mistaken for physical forecasts.
• Boundary risk: variance behavior near zero can affect simulation, PDE, and Greek stability.
• Objective risk: one weighting scheme can improve selected quotes while worsening others.
• Recalibration risk: parameter changes can alter attribution and hedges even when market prices move modestly.
• Model-form risk: multiple models can fit current vanilla prices but disagree on path-dependent values.

Common Mistakes

• Saying the model makes volatility mean-reverting when the specified state variable is variance.
• Including the expected asset return (\mu) as a parameter to calibrate from option prices under the risk-neutral pricing equation.
• Claiming Heston has no closed-form European-option solution.
• Treating one fitted parameter set as permanent.
• Calibrating to implied volatilities without stating weights, quote filters, and forward inputs.
• Ignoring numerical convergence because the displayed price looks precise.
• Assuming a lower calibration error automatically produces better hedges or risk estimates.
• Treating (\sqrt{\theta}) as exactly equal to expected future volatility.
• Reading the Feller condition as a complete model-approval test.
• Using a naive Euler scheme without checking negative variance steps and discretization bias.
• Comparing parameter sets calibrated with different quote filters, objectives, or market inputs.
• Recalibrating during P&L attribution without separating market moves from parameter changes.

Authoritative Sources

• Steven L. Heston, A Closed-Form Solution for Options with Stochastic Volatility
• Federal Reserve, The Pricing Kernel in Options
• Federal Reserve, Backtesting Market Risk Models
• Roger Lord, Remmert Koekkoek, and Dick van Dijk, A Comparison of Biased Simulation Schemes for Stochastic Volatility Models

Related Terms

• Option Pricing Models: The broader family of valuation frameworks and numerical methods.
• Black-Scholes Option Pricing Model: The constant-volatility benchmark used for comparison and implied-volatility quotes.
• Implied Volatility: The model input backed out from an option price.
• Volatility Surface: The strike-and-maturity data commonly used for calibration.
• Option Greeks: The local and higher-order sensitivities used in model risk reports.
• Exotic Option: A nonstandard payoff whose value can depend strongly on stochastic-volatility dynamics.

FAQs

Check Your Understanding

For an actual valuation, use current market data, product terms, model documentation, calibration records, validation findings, and executable price evidence. This article is for financial education only and is not personalized investment, derivatives, valuation, legal, accounting, or tax advice.