Parametric Fits: SABR
Objectives
- State the SABR stochastic-volatility dynamics and name what each of its four parameters — alpha, beta, rho, nu — controls on the smile.
- Map a fitted smile's level, slope, and curvature onto alpha, rho, and nu, and explain why beta is usually fixed rather than estimated.
- Read a SVQ SABR calibration critically — judging when the fitted parameters are genuine evidence about the move the market is pricing, and when they are an over-fit to noise.
In the last chapter you saw the volatility smile as raw data: a scatter of implied volatilities, one per strike, refusing to lie flat the way Black-Scholes promised. That scatter is honest, but it is also unusable as it stands. You cannot interpolate a price between two listed strikes from a cloud of dots, you cannot extrapolate into the wings where no options trade, and you certainly cannot compare today's smile to yesterday's by squinting at two charts. You need a curve — a small handful of numbers that reproduces the whole smile and that you can track through time. That is what a parametric fit gives you, and SABR is the parametric fit the industry settled on.
Let me give you the intuition before a single Greek letter.
Black-Scholes assumes volatility is a constant. The market plainly disagrees — that is the smile. The natural repair is to let volatility itself be random: to say the asset's volatility is its own little process, wandering up and down, sometimes correlated with the asset's own moves. SABR is the cleanest possible version of that idea. The name is an acronym for Stochastic Alpha Beta Rho, and it describes two coupled random walks: one for the forward price of the asset, and one for that forward's volatility. The genius of the model is not its realism — it is deliberately simple — but that someone derived a closed-form formula turning its four parameters directly into a Black-Scholes implied-volatility smile. You fit four numbers; you get the whole curve.
Now the dynamics. Let the forward price be F and its volatility be a quantity we call alpha. SABR says the forward moves by an amount proportional to alpha times F raised to a power beta, driven by random noise. Meanwhile alpha itself moves randomly, scaled by a parameter nu. And the two sources of randomness — the one shoving the price and the one shoving the volatility — are correlated, with correlation rho. That is the entire model. Four parameters: alpha, the overall level of volatility; beta, the backbone that fixes how volatility scales with price; nu, the volatility of volatility; and rho, the correlation between price and vol.
The reason SABR earns its keep is that each parameter owns a recognizable feature of the smile, so fitting it is not a black box — it is reading a face.
Alpha sets the height. Crank alpha up and the entire smile lifts; it is, near the money, essentially the at-the-money implied volatility. When you ask "how much does the market think this thing moves," alpha is the first-order answer.
Rho sets the tilt. Equity smiles slope down to the right — out-of-the-money puts cost more than calls — because crashes and volatility arrive together. SABR encodes that with a negative rho: price down, vol up. The more negative rho is, the steeper the skew. Rho is the parameter that tells you the move the market fears is asymmetric.
Nu sets the curvature — the smile. A large vol-of-vol fans the wings upward, making far out-of-the-money options on both sides expensive, because if volatility can lurch, extreme strikes have a real chance of paying off. Nu is the market's price on surprise.
Beta, the backbone, is the subtle one, and here is the practical truth: you almost never fit beta. Beta and rho fight over the same feature — both influence the skew — so estimating them jointly is unstable; the calibration cannot tell them apart. So you fix beta from a prior view of the asset class and let the other three do the work. Beta near one says volatility is roughly proportional to price, the lognormal world equities tend to inhabit. Beta near zero says volatility is constant in absolute terms, the normal world that rates desks use, and the only setting that handles negative forwards. Beta of one-half sits between. Fix it; calibrate alpha, rho, and nu against it.
SVQ's SABR calibration panel shows four fields for each liquid expiry — alpha, beta, rho, nu — with beta pinned (SVQ defaults to one for equity-style underlyings) and the other three solved to minimize the gap between the model smile and the listed implied vols. Together they read as a sentence. Suppose SVQ returns alpha around zero-point-two-eight, rho around minus-zero-point-six, and nu around one-point-one. That reads: "the market is pricing roughly twenty-eight percent at-the-money volatility, a strong fear-driven downside skew, and a high price on surprise." What separates an analyst from a button-pusher is the residuals SVQ plots beneath the fit. When three liquid strikes hug the curve and the fitted rho is stable from one expiry to the next, that minus-zero-point-six is genuine evidence about the asymmetric move being priced. If the curve only matches because two illiquid wing quotes are dragging it, the parameters are fitting noise, and the story they tell is fiction.
A warning that belongs at Tier 300. SABR's formula is an asymptotic approximation — accurate for short-to-medium expiries near the money, but it can misbehave for very long dated options or very low strikes, where the classic version can even produce negative probabilities. SVQ guards against this, but you should never treat a fitted smile as gospel in the deep wings where the approximation strains and where, conveniently, there is no liquidity to check it against either. The parameters are a summary, not an oracle.