Exotic Options

Vanilla calls and puts depend only on the terminal spot $S_T$. Exotic options depend on more, the path the underlying took, or the payoff is a non-smooth function of the terminal price. That extra structure is where the interesting mathematics and the real hedging danger live: some exotics have elegant closed forms, others force numerical methods, and many have Greeks that blow up in ways no vanilla does. Understanding which is which, and why, is the core of the exotics desk.

The main families

• Barrier options. The payoff activates (knock-in) or extinguishes (knock-out) if the spot touches a barrier $B$ before expiry. A down-and-out call, for instance, is a call that dies if $S$ ever falls to $B$. Path-dependent, but only through the running extremum.
• Asian options. The payoff uses the average price over the life, e.g. $(\bar S - K)^+$ with $\bar S = \tfrac1n\sum S_{t_i}$. Averaging dampens terminal manipulation and lowers vol, so Asians are cheaper than vanillas, popular in FX and commodities for hedging average exposures.
• Digitals (binary options). Pay a fixed amount if a condition holds: a cash-or-nothing call pays $1 if $S_T > K$, else 0. Discontinuous payoff.
• Lookbacks. Pay off on the maximum or minimum over the life, e.g. $S_T - \min_t S_t$ (buy at the low). Strongly path-dependent and expensive.

Digitals: a closed form and a hedging nightmare

A cash-or-nothing digital call paying $1 if $S_T > K$ is priced directly by risk-neutral expectation:

$V_{\text{digital}} = e^{-rT}\,\mathbb{Q}(S_T > K) = e^{-rT}N(d_2).$

Equivalently, and this is the load-bearing insight, a digital is the strike-derivative of a vanilla call:

$V_{\text{digital}} = -\frac{\partial C}{\partial K},$

so a digital is replicated by an infinitely tight call spread: long a call at $K-\varepsilon$, short at $K+\varepsilon$, scaled by $1/2\varepsilon$. This is exactly how desks hedge and price them, and it immediately exposes the danger: near expiry with spot near the strike, the digital's delta and gamma explode (the payoff approaches a step function, whose derivative is a spike). No finite hedge holds; traders instead over-hedge with a wider call spread (a conservative "digital = call spread" super-replication) and eat the bid-offer. Digitals are also acutely skew-sensitive: because they price $\partial C/\partial K$, they depend directly on the smile slope, not just the ATM vol.

Barriers: closed forms via the reflection principle

Under Black-Scholes with constant vol, many barriers have closed forms thanks to the reflection principle for Brownian motion: the probability of a path hitting a level and ending somewhere is computed by reflecting the path across the barrier. The clean result is the in-out parity:

$C_{\text{knock-in}} + C_{\text{knock-out}} = C_{\text{vanilla}},$

because owning both the in and the out versions guarantees you own a live call whether or not the barrier is hit. For a down-and-in call with barrier $B < K$, the reflection principle gives an explicit formula involving $(B/S)^{2\lambda}$ terms with $\lambda = (r - q + \tfrac12\sigma^2)/\sigma^2$, an image charge at the reflected spot. But these formulas assume constant vol; with a real smile, the barrier's sensitivity to smile dynamics near $B$ means local or stochastic vol is essential, and the pricing goes numerical.

Asians and lookbacks: why they need Monte Carlo or PDEs

Arithmetic Asians have no closed form. The payoff depends on $\bar S = \tfrac1n\sum S_{t_i}$, a sum of correlated lognormals, which is not lognormal, the integral has no elementary evaluation. Options:

• Monte Carlo, by Feynman-Kac, the price is $e^{-rT}\mathbb{E}^{\mathbb{Q}}[\text{payoff}]$; simulate paths, average the payoff along each, discount. This is the workhorse for path-dependent and high-dimensional exotics, with variance reduction (the geometric-average Asian does have a closed form and makes an excellent control variate).
• PDE with an augmented state. Add the running average (or running max, for barriers/lookbacks) as an extra state variable, turning the 1-D Black-Scholes PDE into a 2-D PDE solved by finite differences. This works when the path-functional is Markov in one extra dimension.
• Moment matching (Turnbull-Wakeman): approximate the arithmetic average as lognormal by matching its first two moments, fast and often accurate enough for pricing.

The general rule from Feynman-Kac: low-dimensional and Markov $\to$ PDE; high-dimensional or complex path-dependence $\to$ Monte Carlo.

Worked example: down-and-out vs vanilla

Consider a 1-year call, $S = 100$, $K = 100$, $\sigma = 20\%$, $r = 0$, vanilla worth $\approx 7.97$. Add a knock-out barrier at $B = 90$. By in-out parity the knock-out call is $C_{\text{vanilla}} - C_{\text{knock-in}}$; since the down-and-in call requires first falling to 90 then finishing above 100, it captures only a slice of the value, so the knock-out is cheaper than the vanilla, you paid less because you gave up all paths that dip to 90. The closer the barrier to spot, the cheaper the knock-out and the larger its negative gamma near the barrier: as spot approaches 90, the option's value gaps toward zero, so delta swings violently, the classic barrier hedging hazard, where market makers can get pinned defending a barrier.

What breaks in practice

• Discontinuous Greeks. Digitals near strike and barriers near the barrier have delta/gamma that spike or flip sign; delta-hedging is unstable and desks super-replicate with spreads, accepting a bid-offer buffer. Pin risk and barrier "defense" are real P&L events.
• Model risk dominates. Barriers and forward-starting exotics depend on smile dynamics, which local vol gets wrong; the price can differ 10–30% across local-vol, stochastic-vol, and LSV models even when all fit vanillas, the exotic is a bet on the model, not just the market.
• Discrete monitoring. Textbook barriers assume continuous monitoring; real contracts check the barrier daily or at fixes. Discrete monitoring makes a knock-out worth more (harder to breach), and the Broadie-Glasserman-Kou continuity correction shifts the effective barrier by $\approx 0.58\,\sigma\sqrt{\Delta t}$.
• Monte Carlo cost and bias. Path simulation is slow, has discretization bias (finer time steps for barriers), and Greeks by naive bumping are noisy, requiring pathwise/adjoint (AAD) sensitivities.

In interviews

Know the taxonomy, barriers (knock-in/out, path-dependent via the extremum), Asians (average, cheaper, no closed form), digitals (binary, $e^{-rT}N(d_2)$), lookbacks (max/min). The two derivations they love: digital = $-\partial C/\partial K$ = tight call spread, priced $e^{-rT}N(d_2)$, with exploding gamma near strike/expiry and heavy skew sensitivity; and in-out parity $C_{\text{in}} + C_{\text{out}} = C_{\text{vanilla}}$ with barrier closed forms from the reflection principle. Explain the numerical fork via Feynman-Kac: PDE for low-dimensional/Markov, Monte Carlo for path-dependent/high-dimensional, with the geometric-Asian closed form as a control variate. The mature point: exotics are model-dependent, their price is a bet on smile dynamics, so stochastic or local-stochastic vol matters far more than for vanillas.