The Volatility Surface: Reading the Model's Own Errors

The pricing formula maps volatility to price and is monotonic in it: more volatility always means a more valuable option. A monotonic function can be inverted, so given a market price you can solve for the volatility that reproduces it.

That number is the implied volatility, and it is how options are quoted and compared in practice. Traders discuss vol rather than premium, and for good reason.

A premium is not comparable across contracts. A 30 call and a 5 call may be identical value propositions differing only in strike, spot and maturity. Converting to implied volatility divides all of that out, leaving one number on a common scale.

def implied_vol(price, S, K, r, T, kind="call", lo=1e-6, hi=5.0):
    for _ in range(200):                      # bisection: robust, monotone
        mid = (lo + hi) / 2
        if black_scholes(S, K, r, mid, T, kind) < price:
            lo = mid
        else:
            hi = mid
    return (lo + hi) / 2

implied_vol(9.925, 100, 100, 0.04, 1.0)    # about 0.200, recovering the input

Bisection rather than a faster method, because monotonicity guarantees it converges and nothing guarantees a good starting guess.

Now the test. In the model, volatility is a property of the underlying asset: a single parameter describing how it moves.

That has a sharp implication. Every option on the same underlying with the same expiry must imply the same volatility, regardless of strike. Plot implied volatility against strike and you should get a horizontal line.

You do not. In equity index options, implied volatility falls steadily as the strike rises: low-strike puts imply substantially higher volatility than high-strike calls. The shape is called the skew, or in markets where both wings are elevated, the smile.

This is not a small residual or a data artefact. It is large, it is persistent across decades and markets, and it is the single clearest evidence that the model's assumptions are wrong.

The productive response is not to discard the model. It is to treat the shape as information: the market is telling you, in the model's own units, exactly how its beliefs differ from the model's.

Translate the shape back into a statement about the distribution of future prices.

The model assumes returns are normally distributed, with no jumps and constant volatility. A high implied volatility on a low-strike put means the market is paying more for that option than the model says it is worth. Since the option only pays off after a large fall, the market is assigning more probability to large falls than the model does.

So the skew says the real distribution has a fat left tail. Which is exactly what equity index returns are observed to have.

Two mechanisms in the underlying produce this and both are real. Volatility is not constant: it rises sharply during declines, so a fall tends to be followed by more movement rather than less. And prices jump on news, which no continuous-rebalancing argument can hedge.

The skew is therefore not a mispricing to be arbitraged. It is the market correcting a known deficiency of the quoting convention, in the only language the convention provides.

The last box is a stronger result than the skew itself. The full set of option prices across strikes does not merely hint at the market's distribution: it determines it.

Mark Rubinstein set out a construction for this in "Implied Binomial Trees", Journal of Finance, volume 49, issue 3, 1994, pages 771 to 818, inferring risk-neutral probabilities from simultaneously observed option prices and building a recombining binomial tree consistent with them.

The consequence is that an option market is a probability forecast in disguise. Rather than a single expected value, it publishes an entire distribution over outcomes, updated continuously, with real money behind every point on it.

One caveat carries over from the hedging lesson and it is essential: what is recovered is the risk-neutral distribution. It is not what the market expects to happen.

Strike is one dimension. Maturity is the other, and together they make the surface.

Implied volatility varies with time to expiry in ways that carry information about timing rather than magnitude. When short-dated volatility sits above long-dated, the market expects near-term turbulence that it does not expect to persist, which is the usual pattern during a shock. The reverse, with long-dated above short, is the typical calm-market shape: nothing is happening now, and uncertainty accumulates with horizon.

Scheduled events show up crisply. An expiry that falls after an announcement carries visibly higher implied volatility than one falling just before, and the gap is an estimate of the move the market is pricing for that event.

This is directly tradable in principle and it is why the previous lesson's warning matters. Buying that elevated expiry means buying the event, and the elevation disappears the moment the event resolves, whichever way it goes.

The most consequential misreading in this subject is treating implied volatility as the market's estimate of future volatility.

It is a risk-neutral quantity, so it contains a forecast plus whatever premium sellers demand for bearing the risk. Those are different things and they do not have to be close.

The reason a premium exists is structural. Selling options means accepting many small gains and rare large losses, and those losses arrive precisely when everything else in a portfolio is also losing. Insurance that fails in the exact scenario you bought it for is worth less, so protection that pays out in a crisis commands a price above its actuarial value. That gap is the variance risk premium, and it is a well documented feature of index option markets rather than an anomaly.

Two practical consequences. Implied volatility tends to sit above subsequently realised volatility more often than not, which is not a free trade because the losses are concentrated and large. And using implied volatility as a forecast input to a risk model imports the premium as though it were an expectation.

For anyone hedging, the surface's dynamics matter as much as its shape, because they determine what happens to a position when the underlying moves.

Two conventions bracket the possibilities. Under sticky strike, each strike keeps its implied volatility as spot moves, so an option's vol is fixed and only its moneyness changes. Under sticky delta, or sticky moneyness, the shape travels with spot, so an option's implied volatility changes as it becomes more or less in the money.

The choice is not cosmetic: it changes the computed delta. Under sticky delta, a spot move implies a volatility move, which implies an additional value change through vega, and that has to be added to the delta from the previous lesson.

A desk that hedges with model deltas while the market actually behaves according to the other convention will be systematically mis-hedged, in the same direction, every day. The error is small per rebalance and does not cancel, which is the worst combination.

The habit worth building is to read the surface as a set of claims and ask which you disagree with. A view that the market overprices crash risk is a view about the slope. A view that a specific event will be quieter than expected is a view about one expiry. Each is expressible as a position, and each is distinct from a view on direction.

That is the final answer to what an option trader trades. Not the asset, and not even volatility as a single number, but the shape of the market's distribution, one region at a time.

Four lessons, and the arc is one idea developed to its conclusion.

At expiry, an option is arithmetic. Before expiry, put-call parity ties calls to puts with no model at all. Pricing either one requires a model, and the useful one works by replication, which cancels direction and leaves volatility. The greeks then decompose what a replicated position is still exposed to. And the surface is the market's correction to the model, expressed in the model's own units.

The transferable idea is the last one. A wrong model that everyone agrees on is more useful than a right model nobody shares, because agreement makes deviations measurable. Black-Scholes survives not despite being violated but because the violations are legible.

What this does not cover is what happens when the hedge fails at scale: sizing a position so a gap does not end the business. That belongs to risk management, and it is a discipline rather than a formula.