SABR Model: Alpha, Beta, Rho, Nu, Calibration, and Model Risk

For educational purposes only; not investment advice.

Direct answer

The SABR model is a stochastic-volatility model in which a forward F_t and its volatility level α_t move randomly and their shocks are correlated. It is widely used to parameterize an implied-volatility smile with a small, interpretable parameter set and to generate smile-consistent prices and sensitivities.

SABR is not one universal volatility surface. It is commonly calibrated separately by expiration or rate-option expiry/tenor, using a Black, normal, or shifted quotation convention. The familiar Hagan implied-volatility formula is an asymptotic approximation to the model, not an exact price and not automatically arbitrage-free at all strikes and maturities.

Dynamics and parameter roles

Under a pricing measure, the classic forward SABR dynamics are:

dF_t = α_t F_t^β dW_t¹

dα_t = ν α_t dW_t²

d⟨W¹,W²⟩_t = ρ dt

with current α > 0, commonly 0 ≤ β ≤ 1, ν ≥ 0, and −1 < ρ < 1.

Alpha (α) sets the current volatility scale, but is not generally equal to ATM implied volatility.
Beta (β) controls elasticity: β = 1 is lognormal-like, β = 0 is normal-like, and intermediate values create CEV scaling.
Rho (ρ) correlates forward and volatility shocks and strongly influences smile asymmetry.
Nu (ν) is volatility of volatility and commonly increases smile curvature as it rises.

These interpretations are not independent. Alpha and Beta can compensate around ATM; Rho and Nu can compensate across a limited strike range. Many implementations fix or tightly constrain Beta, then calibrate Alpha, Rho, and Nu. A low fitting error does not prove parameter identification or realistic future smile dynamics.

For Black-style SABR, a leading ATM intuition is:

ATM Black IV ≈ α / F^(1−β)

with maturity corrections involving all parameters. Normal SABR is often used where rates or forwards can be near or below zero. Shifted SABR applies a displacement to F and K; that shift changes dynamics and tails and must be governed as a model parameter.

Parameter-scale example

Let forward F = 100, maturity T = 1 year, fixed β = 0.50, and α = 2.00. The leading ATM approximation is:

ATM IV ≈ 2.00 / 100^(1−0.50) = 2.00 / 10 = 20.00%

If the forward falls to 81 while Alpha is held at 2.00, the same approximation becomes:

2.00 / 81^0.50 = 2.00 / 9 = 22.22%

Thus unchanged Alpha does not imply unchanged ATM IV when β < 1. Conversely, comparing Alpha across assets with different units or Beta values is generally meaningless.

Now compare two illustrative configurations with the same F, T, β, and α:

Set	`ρ`	`ν`	Typical qualitative effect
A	`−0.20`	`0.30`	milder negative skew and curvature
B	`−0.70`	`0.80`	steeper downside skew and stronger curvature

That description is conditional, not a numerical price guarantee. A valid comparison must evaluate the selected Black/normal formula at actual strikes, convert IV errors into price errors, and inspect extrapolation. Parameters B can fit observed nodes yet produce unstable or implausible far wings.

Calibration and validation checklist

Fix underlying, forward construction, discounting, expiry, tenor, settlement, and quote convention.
Clean same-time Bid/Ask quotes and retain liquidity weights and exclusion reasons.
Decide explicitly among Black, normal, and shifted SABR; never mix their volatility units.
Fix or bound Beta with documented rationale and repeat valuation under alternative Beta values.
Enforce α > 0, ν ≥ 0, and a safe interior bound for Rho.
Use multiple initial guesses and inspect local minima, parameter covariance, and day-to-day jumps.
Minimize price or Vega-weighted errors as appropriate, not unweighted IV error by habit.
Report fit separately for ATM, downside, upside, and Bid/Ask coverage.
Perform leave-one-out or out-of-sample checks instead of relying only on in-sample residuals.
Convert the fitted smile back to option prices and check strike monotonicity and convexity.
Inspect calendar consistency across separately calibrated expirations.
Compare Hagan approximation prices and Greeks with a more accurate numerical method at long maturities, high Nu, extreme Rho, and distant strikes.
Stress Alpha, Beta, Rho, Nu, forward, shift, rates, and correlations jointly.
Preserve market data, versioned formula, objective function, bounds, solver, initial values, and calibration timestamp.

Common misconceptions

“SABR is an exact closed-form pricing model.” The popular formula is an asymptotic implied-volatility approximation.
“Alpha is ATM IV.” It also depends on forward scale, Beta, maturity, Rho, and Nu corrections.
“Rho is historical price/volatility correlation.” It is a risk-neutral model parameter calibrated to option prices.
“Nu controls only the wings.” Parameters interact, and changing Nu can affect ATM and hedge sensitivities.
“Beta can always be estimated freely.” Alpha-Beta collinearity often makes the fit unstable.
“A perfect smile fit proves good hedges.” Different dynamics can fit the same vanilla snapshot and hedge exotics differently.
“Normal and Black volatilities are interchangeable percentages.” They use different units and pricing conventions.
“Shifted SABR solves negative rates without cost.” The displacement is an additional structural assumption affecting tails.
“Positive fitted IV means no arbitrage.” The resulting price curve can still violate convexity or calendar conditions.
“One SABR fit describes the entire surface.” Independent slices can be inconsistent across expirations and tenors.