ZABR from zero
1/5SABR with a flexible backbone
SABR says the forward diffuses with vol proportional to Fᵝ -- a power law. That one exponent β controls the entire backbone. ZABR replaces the power law with a general functionγ(F). Same SABR structure, but the backbone can take any shape.
In standard SABR, the forward SDE is:
ZABR generalizes this by replacing Fᵝ with an arbitrary smooth function γ(F):
SABR gives you one bendable rod: the power law Fᵝ. Changing β bends the rod one way or another, but it is always the same shape family. ZABR lets you swap in an entirely different rod before you start bending. The rod's shape is the backbone, and ZABR says: choose whatever shape fits your market.
If you set γ(F) = Fᵝ, you recover SABR exactly. ZABR is a strict generalization. The question is: when does the extra flexibility matter?
The γ function
In SABR, γ(F) = Fᵝ. In ZABR, γ(F) can be piecewise, spline, or any smooth positive function. This means the local vol backbone can have kinks, shelves, inflections -- shapes that no single power law can produce.
The backbone function γ(F) tells the model: for each level of the forward, how sensitive is the local vol to stochastic vol shocks? A high γ(F) at a particular level means vol is very reactive to price being at that level. A low γ(F) means vol is muted there.
SABR's Fᵝ: A monotonic function. When β < 1, γ grows sublinearly -- vol sensitivity is relatively higher at low F. When β = 1,γ grows linearly. But it is always smooth, monotonic, and concave.
ZABR's general γ(F): Can be non-monotonic. Can have a shelf (vol sensitivity saturates at low F). Can have a kink (abrupt change in sensitivity at some price level). Can be piecewise linear, spline, or any parametric form you choose.
Drag the β slider and compare SABR's power-law backbone with the two ZABR alternatives. The "shelf" backbone flattens at low F -- it says vol sensitivity saturates when the forward is very low, which prevents the explosion that SABR with low β produces near zero. The "S-curve" backbone concentrates vol sensitivity in a band around the current forward, which is a different structural assumption about how markets behave.
The designer above lets you drag control points to create any backbone shape and see the resulting smile. The connection between backbone shape and smile shape is direct: where γ(F) is steep, the smile has more curvature; where γ(F) is flat, the smile is smoother.
Why generalize the backbone?
Some markets have smiles that SABR's Fᵝ cannot match. When the backbone itself is wrong, no amount of parameter tweaking can fully save the fit. ZABR lets the backbone adapt.
Rates near zero. When interest rates are near zero or negative, the SABR backbone creates problems. With lowβ, the Fᵝ term can produce extreme vol at low F, creating unrealistic smiles. With high β, the model cannot handle negative rates at all. ZABR with a backbone likeγ(F) = (F + d)ᵝ (shifted power law) or a tanh function handles this gracefully.
Credit spreads. CDS option smiles often have shapes that SABR misses systematically in the left wing. The spread dynamics at low levels (near default) behave differently than at high levels. A piecewise backbone can capture this transition.
Equity vol during regime changes. After a large selloff, the smile can develop features (kinks, extra steepness in specific strike ranges) that SABR's smooth power law cannot reproduce. ZABR with a spline backbone can capture these transient features.
Switch between the two presets above. In the "normal market" case, SABR and ZABR produce nearly identical smiles -- the extra flexibility of ZABR is not needed. In the "kinked left wing" case, SABR systematically misses the kink. ZABR's backbone can adapt to match it.
The lesson: ZABR earns its keep only when there is a systematic backbone misfit. If SABR fits well, there is no reason to add the complexity of a custom backbone. The model selection criterion is empirical: does the residual between SABR's best fit and the market show a pattern that a different backbone could fix?
The asymptotic expansion
ZABR uses the same Hagan-style asymptotic expansion as SABR, but withγ(F) replacing Fᵝ. The formula structure is identical; only the backbone function changes.
The Hagan-Woodward SABR formula (2002) is an asymptotic expansion of the implied vol in powers of the vol-of-vol ν and the expiry T. The key building block is a mapping from the forward level to a "normal vol" space via an integral involving the backbone:
The rest of the Hagan formula -- the z-to-x mapping, the correction terms -- is structurally the same. You replace every instance of Fᵝ with γ(F) and every instance of the backbone integral with its numerical value. The expansion remains valid to the same order.
Why this matters: The asymptotic expansion is fast. For each (K, T) pair, you evaluate one integral (numerically), plug into the same Hagan-style formula, and get an implied vol. No PDE, no Monte Carlo. This is what makes ZABR practical: it has the speed of an asymptotic formula with the flexibility of a custom backbone.
Accuracy limitations: The Hagan expansion is only first-order in T. For long-dated options, it can be inaccurate. This is the same limitation as SABR itself -- the expansion is for short-to-medium expiries. For long dates, you need a PDE solver or Monte Carlo, regardless of whether you use SABR or ZABR.
Alternative: PDE approach. Instead of the asymptotic expansion, you can solve the ZABR pricing PDE directly. This is more accurate but slower. Some implementations use the asymptotic expansion as a first guess and refine with a PDE correction.
ZABR in practice
ZABR is a specialist tool. It is used by rates desks for negative-rate environments and by exotics desks where backbone misfit causes hedging errors. It is less common than shifted SABR, which is simpler and often good enough.
Rates markets: The primary user base. When rates went negative in EUR and JPY, desks needed models that could handle F < 0. Shifted SABR (with γ(F) = (F + d)ᵝ) was the quick fix. Full ZABR with a custom backbone was the higher-end solution for desks that needed more precise wing fitting.
Exotics pricing: Path-dependent products (CMS caps, range accruals) are sensitive to the backbone shape because the payoff depends on how the forward moves through different levels. A wrong backbone means wrong dynamics, which means wrong exotic prices even if the vanilla smile fits. ZABR addresses this by letting the backbone match the empirical dynamics.
Calibration: Fitting γ(F) to market data is harder than fitting β alone. With SABR, you optimize over four parameters. With ZABR, you optimize over the parameters ofγ (which could be a spline with many knots) plusα, ν, and ρ. This is a higher-dimensional problem that requires more data and more careful regularization.
When not to use ZABR:
1. When SABR fits well. Extra complexity without extra value is just extra risk. If the SABR residuals are small and structureless, keep it simple.
2. When you do not have enough data to constrain the backbone. A flexible γ with sparse data leads to overfitting. You need enough liquid strikes across the smile to justify the extra degrees of freedom.
3. For crypto vol surfaces. Crypto desks typically use SVI/SSVI for static fitting and do not need the dynamic backbone story that ZABR provides. The smile shapes are better handled by direct parameterizations than by modifying a stochastic vol backbone.
Black-Scholes (γ = F, no stochastic vol) →SABR (γ = Fᵝ, stochastic vol) →ZABR (γ = general function, stochastic vol). Each step adds flexibility and complexity. Use the simplest model that fits your market and supports your hedging needs.
Where to go next:
SABR Model -- the foundation that ZABR generalizes
Displaced Diffusion -- the simplest shift approach
Stochastic Local Vol -- an alternative approach to backbone flexibility
Heston Model -- stochastic vol with a different variance process