The Volatility Smile: Why Implied Vol is a Curve, Not a Number

Walk into any derivatives desk on a Monday morning, open a Bloomberg terminal, and pull up the option chain on the SPX. You’ll see a column labeled “IV” or “Imp Vol” with a number for each strike. ATMs trade around 16%. Strikes $5\%$ OTM trade at $21\%$. Deep OTM puts trade at $30\%$, $40\%$, even higher in a crisis.

If Black-Scholes were true, that column would contain a single number for every strike. Markets do not trade that way. The pattern of IV across strikes, the smile or skew, reflects how market makers price large moves.

What the Smile Looks Like

Equity index options have a negative skew. Puts trade richer than calls. Pull up SPX and you’ll see something like:

| Strike | IV | | | | | 3800 (deep OTM put) | 32% | | 4200 (slight OTM put) | 22% | | 4500 (ATM) | 16% | | 4800 (slight OTM call) | 17% | | 5100 (deep OTM call) | 19% |

This pattern is characteristic of equity indexes. The slope is negative across strikes, with the strongest curvature in the puts. Single-name equities have a similar but less extreme skew. FX options have a symmetric smile, with both wings elevated. Interest-rate caps have a smile with a tilt, with different levels on the two sides because rate distributions are asymmetric.

Why the Smile Exists

The single best diagnostic is to invert observed option prices through the Black-Scholes formula and ask: “If I assumed lognormal returns, what vol would I have to plug in to reproduce this market price?” The number you get is the implied volatility. The fact that it depends on the strike is exactly the smile.

If you believe lognormality, the only IV curve is flat. Markets are telling us that returns have fatter tails. A lognormal distribution assigns about $5\%$ probability mass to a $3\sigma$ move; markets price that as roughly $3\%$ but imply about $8\%$. The difference is the IV gap.

There are three stylized facts that any decent model of the smile has to reproduce (Hull, 2018, Ch. 18):

1. The skew is steep on the downside. Traders pay up for downside protection. The single most heavily traded options contract in the world is the 1-month 5-delta put on the S&P 500. People hedge crashes.

2. The smile is roughly flat for short maturities. Very short-dated options, those with less than a week to expiry, do not show much smile in calm markets. Vol stays close to realized volatility because the central limit theorem has had little time to average out the path; each tick matters.

3. The smile flattens and shifts for longer maturities. As $T$ increases, the smile becomes less pronounced relative to its level, and the smile itself becomes more symmetric. This is roughly consistent with mixed diffusion-jump models (Merton, 1976) where jumps dominate the short end and diffusion governs the long end.

Three Mechanisms Behind the Smile

Several economic mechanisms can generate the smile. Each points to a different modeling choice.

Mechanism 1: Stochastic Volatility

Volatility is not constant. When markets move a lot, vol tends to rise. This is the leverage effect (or “variance feedback” depending on the school of thought). The returns distribution becomes a mixture of lognormals: when vol is high, the realized returns have a wider distribution, so tail probabilities are higher than lognormal.

The mixture flattens and shifts the implied distribution. Quantitatively (Wilmott, 2007, Vol. 1, Ch. 13):

where $p(\sigma)$ is the distribution of the integrated variance. If we Taylor-expand around the average vol $\bar\sigma$ using $f^{\text{BS}}(S, \sigma)$ as a function of $\sigma$:

The second term is vomma. When volatilities are mixed, options are priced near the average vol plus a convexity correction. Convexity increases with moneyness for puts, so the skew is more negative on the put side. This matches the equity-market pattern.

Mechanism 2: Jump Risk

Black Monday in 1987 was a 23% one-day move in the S&P 500. Lognormal with vol 20% assigns that event probability $\sim 10^{-30}$. The market price of a 23% jump is much higher than the lognormal probability.

Merton (1976) modeled this by adding a Poisson jump component to the stock SDE:

where $dJ_t$ is a compound Poisson process. The distribution of returns is now a mixture of diffusion and jumps, with heavier tails than either component alone.

A pure jump model produces a flat smile but elevated wings relative to lognormal. A pure stochastic vol model produces a symmetric smile. The equity skew needs both: stochastic vol for the smile shape and jumps (or a left-skewed vol process like Heston with negative correlation) for the asymmetry.

Mechanism 3: Crashophobia and Hedging Pressure

Pure economics isn’t the full story. The skew is partly demand-driven: investors buy puts, dealer hedges depress the implied vol of puts and inflate that of calls (synthetic underhedging). In stressed markets, demand for out-of-the-money puts spikes even more, and the skew turns into a tear. IV on 5-delta puts can run $40\%$ while 25-delta puts trade $20\%$.

This is a behavioral mechanism: the market is not pricing tail risk at the true objective probability but at a higher subjective fear-weighted probability. After 1987, the equity skew became permanently steeper; after 2008 and 2020, the same thing happened. Traders call this effect “fear premium” (Hull, 2018, Ch. 18.7).

The Math: Tighter Than the Intuition

Quantify the smile in terms of risk-neutral moments. The price of a digital or out-of-the-money option is sensitive to the local density. Specifically (Wilmott, 2007, Vol. 1, Ch. 13.3):

For a call, the Breeden-Litzenberger (1978) identity gives:

where $q(K)$ is the risk-neutral probability density of $S_T$ at $K$. Equivalently, the IV smile is a smoothed version of the risk-neutral density. ATM IV is the average of the density at the strike; the wings of the smile trace the tails of the density.

If you accept that, then the smile literally is the distribution. The implied vol at $K = 0.85 S_{\text{spot}}$ is the local height of the risk-neutral density at $S_T = 0.85 S_{\text{spot}}$, expressed via a Black-Scholes inversion.

For an index:

• ATM vol ≈ scaling of the entire distribution
• Skew ($d\sigma/dK$ at ATM) ≈ log-volatility-derivative of the spot variance, or roughly the asymmetry of returns
• Curvature ($d^2\sigma/dK^2$ at ATM) ≈ kurtosis of the risk-neutral distribution

This is called the edgeworth expansion of IV around ATM (Bakshi, Chen & Gao, 1995). Quant traders use it all the time to map IV skew into risk-neutral moments.

The SVI Parameterization

You need a closed-form expression for the IV smile for two reasons: first, calibration needs something smooth to fit; second, no-arbitrage conditions need to be enforceable. SVI (Stochastic Volatility Inspired) by Gatheral (2004) is the most popular parameterization because it fits the wings well and has tractable no-arbitrage conditions.

For a slice at time-to-expiry $\tau$, SVI parameterizes the total implied variance $w = \sigma^2 \tau$ as a function of log-moneyness $k = \ln(K/F)$:

Five parameters:

• $a$: the overall level of variance, which sets the ATM vol roughly
• $b$: the slope at the wings; a higher $b$ means a fatter-tailed smile
• $\rho$: the rotation; $\rho = -1$ makes the smile steep on the left, while $\rho = 0$ makes it symmetric
• $m$: shifts the smile left or right, usually negative for equities, so the smile is shifted toward OTM puts
• $\sigma$: the smoothness of the wings

Fitting SVI to market data is well-conditioned in the central region but tricky in the wings. Quants typically fit on a calibration set with weighted least squares, with more weight on the ATM and near-ATM strikes where hedging matters most, and less weight on the far wings where bid-ask is wide.

No-arbitrage conditions on SVI have been studied extensively. Gatheral and Jacod (2018) showed that the local variance must remain non-negative:

These are necessary conditions; sufficient conditions require more work. In practice, no model is arbitrage-free at floating-point precision, so check by computing butterfly spreads:

for any three strikes $K_1 < K_2 < K_3$. If any of these turns negative, you have a calendar-able arbitrage.

What Comes Next

The smile is the market’s read on the distribution. The following articles cover local-volatility extraction, stochastic volatility, and arbitrage-free parametric fitting. Treat the smile as a boundary condition for volatility modeling rather than as a model feature.

What you should take from this article:

1. The smile exists because markets price fatter tails and skew than lognormal.
2. Three sources: stochastic vol, jumps, and demand pressure. None alone matches the equity market; a combination is needed.
3. The skew encodes the risk-neutral distribution quantitatively: its slope reflects asymmetry and its curvature reflects kurtosis.
4. SVI is a common parameterization, and it has no-arbitrage conditions that must be respected.

When you hear someone say “the skew is too steep,” they mean the implied distribution has more left tail than their model can price. When they say “the skew is dislocated,” they mean the whole curve has jumped, which only happens during crashes or vendor data glitches.

SVI in Practice: A Calibration Walk-Through

The SVI parameterization has five parameters per slice: $a, b, \rho, m, \sigma$. In a real calibration, you want smoothness across slices, so neighboring $T$ values should give similar parameter values. The standard approach is:

1. Choose anchor tenors (e.g., 1M, 2M, 3M, 6M, 9M, 1Y, 18M, 2Y).
2. For each tenor, extract the IVs across strikes. Convert to log-moneyness $k = \ln(K/F)$ where $F = S_0 e^{rT}$ is the forward.
3. Compute total variance $w(k, T) = \sigma_{\text{IV}}^2 \cdot T$.
4. Fit SVI parameters to minimize $\sum_i (w_{\text{SVI}}(k_i; a, b, \rho, m, \sigma) - w_{\text{mkt}}(k_i))^2$ weighted by inverse-bid-ask.

Typical fit: 5-10bps RMS error across the strip. Some tenors (especially very near or far) have worse fits; this is normal because the data is sparse or noisy there.

Two fit strategies that production teams use:

Slice-by-slice. Fit each tenor independently. May give wiggles across tenors, but each slice is locally optimal.

Joint with regularization. Fit all slices simultaneously, with regularization terms to enforce that $(a, b, \rho, m, \sigma)$ vary smoothly across $T$. Adds a parameter: the smoothness weight. Production teams tune this to balance fit quality and inter-slice smoothness.

Beyond SVI: The Dynamic Smile

SVI is a static snapshot. For trading strategies that hold positions over weeks, the change in the surface is what matters. The structure of that change is called the smile dynamics.

A simple empirical model: each SVI parameter follows a random walk with drift. Calibrate today; tomorrow’s parameters are yesterday’s plus a small shock. Estimate the covariance of these shocks from historical smile data. This is a “smile covariance” matrix or vol surface dynamics model.

One surprising empirical regularity: the sticky strike vs. sticky delta dichotomy. When the spot moves, do you re-evaluate the surface at the same strike (sticky strike) or the same delta (sticky delta)?

• Sticky strike: the same IV applies to the same strike regardless of spot.
• Sticky delta: the IV at the same delta strike stays the same. As spot moves, the strike moves to maintain delta, and IV is unchanged along that moving line.

Empirically, sticky delta is closer to truth for short-dated options and indices, but for longer-dated options sticky strike is better. Different markets have different dynamics. The idea of mixed sticky or finite sticky parameterizations: there’s a smooth transition between the two regimes.

Practically: a market maker who re-hedges delta every 30 seconds and notices their delta-hedge book has unexplained P&L (their model assumes sticky X but the market is sticky Y) is seeing this effect directly. The fix is to use the right stickiness rule for each market.

The Shape of Different Markets

Each market has a characteristic smile shape that says something about the underlying return distribution.

Equity indices (S&P, Russell). Negative skew, hump-shaped. The OTM put wing rises steeply, while the OTM call wing is only slightly elevated. This indicates fat tails and crash risk. SPX OTM 5-delta puts often trade at $30+\%$ IV versus $16\%$ ATM, roughly twice the at-the-money level.

Single-name equities. Similar to indices but skew is less pronounced. AAPL IV skew flattens about 6 strikes out; SPX skew stays steep across the whole chain. Single-name skew reflects company-specific crash risk (the price collapse risk for individual firms is high).

FX. Symmetric smile, both wings elevated. GBP/USD 5-delta calls trade at similar vol to 5-delta puts. The smile reflects the symmetric tail distribution of FX returns.

Commodities. Often a backwardated term structure with low absolute vol in contango markets. Crude oil smile is roughly symmetric but with a small backward shift reflecting backwardation. Gold smile is more symmetric still. Agricultural commodities have seasonal vol effects.

Rates. Caps and floors on interest rates have an asymmetric smile that’s hard to characterize cleanly. The skew depends on the central bank’s policy regime: when rate changes are biased toward cuts, the floor (downside) skew is steeper.

The Risk-Neutral Density View

Recall Breeden-Litzenberger:

This is the risk-neutral density of $S_T$. The IV smile is a smoothed representation of this density.

The risk-neutral density tells you what the option market thinks the future spot distribution is. The risk-neutral cumulative distribution:

(or via puts). For a normal ($q$ flat), the IV smile is flat. For a left-skewed distribution, the IV smile is steep on the put side.

This gives an interesting quantification: “the market implies 12% probability of $S_T < 1000$ for SPX, when $S_0 = 4500$” is a specific number read off the implied distribution. The objective probability from historical data is much lower, about 3% in the realized distribution. The difference reflects risk premia, hedging demand, and tail overestimation.

Arbitrage Smiles vs. Market Smiles

There’s a subtle distinction. No-arbitrage smiles satisfy all the butterfly and calendar constraints. Market smiles are observed from actual quotes, which might (briefly) violate those constraints.

In practice, observed smiles occasionally have:

• Negative density at extreme strikes: market quotes imply $q_T < 0$ at some strike, which is impossible for a probability density.
• Calendar spreads that are inverted: total variance decreases with maturity for some strikes.
• Vertical arbitrage: $C(K_1) - C(K_2) > S_0 - K_1 e^{-rT}$.

These violations are usually transient. They may appear for an hour and then disappear. The cleanest approach is to maintain a regularized arbitrage-free surface, such as calibrated SSVI, and trade against that surface rather than noisy market quotes.

When the discrepancy is sustained, it may indicate a market-structure issue, such as an unpriced corporate action or a market maker trying to move a position.

The Smile and Realized Vol

A common trading strategy: compare implied vol (the ATM vol of the smile) to recent realized vol. When IV is high relative to realized, the expected return to shorting volatility is positive, and vice versa.

This works in the long run (variance premium is positive historically, especially after 1987) but with substantial variance around the mean. Shorting vol has historically lost money in 2008, 2020, and other tail events, while making slow steady money in between. The strategy is more interesting when timed with vol-of-vol signals: short vol when vol-of-vol is high, long vol when vol-of-vol is low.

The same logic applies to variance risk premium: the historical gap between IV and realized vol is on average 2-4 vol points per year. Traders collect this over time but get hit hard on outlier days.

The Take-Home Points

The volatility smile reflects the market’s view of the future distribution. Stochastic volatility, jumps, and crash-hedging demand combine to produce the observed pattern. SVI is a common parameterization, but it requires careful calibration. Smile dynamics also matter for trading and hedging decisions.

The smile is not a feature of one model. It is a property of the market data that all plausible models must fit. Once you understand that distinction, you can choose a parameterization based on the task rather than treating one model as definitive.

Reading list. Hull, Options, Futures, and Other Derivatives, 10th ed., Ch. 18 (2018). Wilmott, Paul Wilmott on Quantitative Finance, 2nd ed., Vol. 1, Ch. 13 (2007). Gatheral, The Volatility Surface (2006). Bakshi, Chen, Gao, “Empirical Performance of Alternative Option Pricing Models,” Journal of Finance (1997). Merton, “Option Pricing When Underlying Stock Returns are Discontinuous,” JFE (1976).