The Model That Fixed the Hedges: SABR and Managing Smile Risk

The title of this paper is the best thing about it, and it tells you what the authors thought mattered. Not "Pricing Smile Risk." Managing it.

By 2002, the industry had solved the pricing problem. Dupire's local volatility model, and its tree cousins from Derman-Kani and Rubinstein, could reproduce any observed volatility smile exactly. Every desk used them.

Patrick Hagan and his co-authors opened their paper by pointing out that these models, despite fitting today's prices perfectly, produce hedges that are systematically wrong. Not slightly wrong. Wrong in the wrong direction. And since a derivatives desk lives or dies by its hedges, not by its marks, this was a serious accusation.

They were right, and the model they built to fix it, SABR, became the standard for the entire interest-rate options market.

The problem: local volatility predicts the smile will move backwards

Here is the argument, and it is the heart of the paper.

Take a local volatility model calibrated to today's smile. Now ask it a question: if the underlying moves up by one percent tomorrow, what happens to the smile?

Local volatility gives a very specific answer, and it follows from the model's basic structure. Because volatility is a fixed function of the price level, and the smile is currently downward-sloping (higher volatility at lower strikes), the model says: if the underlying rises, it moves into a region of lower local volatility, so the whole smile shifts down and to the left. The smile moves in the opposite direction to the underlying.

The market does the opposite. Empirically, when the underlying moves, the smile tends to move with it: the whole volatility curve slides sideways in the same direction, so that the at-the-money volatility stays roughly where it was relative to the new spot. The market's smile is roughly "sticky delta." Local volatility predicts something close to the reverse.

Now, why does this matter so much? Because your delta, the amount of underlying you hold as a hedge, is not just the sensitivity of the option's price to the spot. It also includes the effect of the spot move on the implied volatility, and therefore on the option's price through vega. If your model has the smile moving the wrong way, your delta absorbs a correction of the wrong sign. You will be hedging with a number that is not just imprecise but pointing the wrong way, and you will bleed money in a manner that is invisible in your daily mark-to-market, because your marks fit the market perfectly. Only the hedging losses accumulate.

This is one of the most important and under-appreciated distinctions in derivatives: fitting today's prices and predicting tomorrow's dynamics are different jobs, and a model can ace the first while failing the second. Hagan and colleagues stated it plainly.

The key idea via analogy: let the forward and the volatility drift together

SABR stands for Stochastic Alpha Beta Rho, which is simply a list of its parameters. The model is deliberately minimal:

• The forward price follows a random process whose volatility is a parameter, alpha, and whose behaviour is shaped by a beta parameter that interpolates between a normal model (where the underlying moves in absolute terms) and a lognormal one (where it moves in percentage terms).
• The volatility itself is a second random process, wandering with its own volatility (nu, the "vol of vol"), which controls the curvature of the smile.
• The two random drivers are correlated by rho, which controls the tilt, the skew.

Notice what is different from local volatility. Here, volatility is genuinely stochastic: it has its own noise, and it is not chained to the price level. So when the forward moves, the volatility does not have to move in the opposite direction to preserve the fit. The smile can, and does, drift along with the underlying. The dynamics come out roughly right.

Four parameters, each with a clean job:

• alpha: the overall level of volatility. Moves the smile up and down.
• beta: the backbone, how at-the-money volatility behaves as the forward moves. Usually fixed by convention rather than fitted.
• rho: the tilt. Negative rho gives you a downward skew.
• nu: the curvature. High vol-of-vol gives a pronounced smile.

The real innovation: an explicit formula

SABR's genuine breakthrough is not the model, which is fairly natural. It is that Hagan and his colleagues found an explicit approximation formula that converts SABR's parameters directly into a Black implied volatility, for any strike and maturity, with no numerical work whatsoever.

That sounds like a technicality. It is the entire reason SABR conquered the market.

Interest rate desks price caps, floors and swaptions across dozens of expiries, dozens of tenors and dozens of strikes, thousands of instruments, updated continuously. A model requiring a numerical PDE solve or a Monte Carlo per instrument is unusable at that scale. SABR gives you the implied volatility from a closed-form expression. You can put it in a spreadsheet. You can put it in every cell of a swaption grid and it recalculates instantly.

The formula is an asymptotic expansion, meaning it is an approximation that gets better as volatility and maturity get small. It is not exact. But it is fast, it is smooth, and it is accurate enough over the range that mattered to the rates market. Practicality won.

Why it mattered

• It is the language of the rates options market. For two decades, swaption and cap volatility surfaces have been quoted, interpolated and risk-managed in SABR parameters. When a rates trader says "the vol of vol on the 5-year-into-10-year is 30," they are speaking SABR.
• It made "smile dynamics" a first-class concern. After this paper, nobody could get away with calibrating a model to today's surface and calling it done. The question "what does your model say happens to the smile when spot moves?" became a standard test, and a lot of models failed it.
• It gave a coherent story for hedging. SABR produces a delta that accounts for the expected co-movement of spot and volatility, and a vega that is meaningful. The corrections that follow (what practitioners call the SABR delta, and the vanna and volga risks) are the daily bread of an options desk.
• It is interpolation as much as it is a model. In practice SABR is often used not as a description of reality but as a well-behaved, arbitrage-friendly way to fill in the gaps between quoted strikes. That is a modest role, and it is the role at which it excels.

The honest limitations

• The formula breaks down at low strikes, and this became a crisis. The asymptotic approximation can produce implied volatilities that imply a negative probability density for low strikes. That is not a rounding error, it is nonsense: the model is asserting that some outcomes are less than impossible. In the low- and negative-rate environment after 2008, when strikes near and below zero became the market's daily reality, this failure went from an academic footnote to a live production problem, and the industry had to scramble. A large literature exists on repairing it, including exact SABR solutions, shifted SABR variants (which move the whole rate axis up so lognormality still works), and the normal SABR case.
• It is an approximation and it degrades with maturity. The expansion is accurate for short maturities and modest volatilities. For long-dated swaptions, where the rates market has plenty of business, the errors grow, and the mispricing can be material.
• It has no mean reversion. SABR's volatility wanders off without being pulled back, which is not how volatility behaves. This makes the model's long-horizon dynamics unrealistic, and it means SABR is a single-expiry model: you fit one set of parameters per expiry and tenor, with nothing linking them. There is no global, term-structure-consistent story. Practitioners fit hundreds of separate SABR models to one surface, which is a candid admission that the model has no view on time.
• The dynamics are better, not right. SABR improves on local volatility's backwards smile, but its predicted smile dynamics are still only approximately what the market does. The delta it produces is better than local volatility's, not correct.
• Beta is unidentifiable from the smile. In practice, beta and rho are nearly interchangeable in their effect on the fitted shape, so beta is almost always fixed by convention (often 0.5, or 1, or 0) rather than estimated. A four-parameter model that is really a three-parameter model.

The one-line takeaway

Hagan and colleagues pointed out that local volatility models, though they fit today's smile perfectly, predict that the smile moves in the wrong direction when spot moves, which corrupts every hedge ratio, and built SABR: a minimal stochastic volatility model with an explicit formula for implied volatility, fast enough to plaster across an entire swaption grid, which became the quoting language of the rates options market despite an approximation that famously falls apart at low strikes.