Pricing Path-Dependent Options

Black–Scholes is a magic trick that only works on one kind of option. It prices a plain European call in closed form — plug in five numbers, out pops a price — and it can do that because the payoff depends on exactly one thing: where the stock lands at expiry. The whole journey in between is irrelevant; only the destination counts. So Black–Scholes never has to know the path. It just needs the final distribution.

But a huge slice of the real derivatives market pays off on the journey, not the destination. An option whose payoff is the average price over six months. An option that pays like a vanilla unless the stock ever touched a knock-out level, in which case it vanishes. For these, the destination tells you almost nothing — you need the entire path. And the moment your payoff needs the path, the clean formula evaporates and you’re left with Monte Carlo, which happens to be built on paths. This lesson is its home turf.

Risk-neutral valuation: pricing in a pretend world

Analogy. Imagine two people betting on a coin. To agree on a fair price for the bet, they don’t need to argue about which way the coin “really” leans — they agree to value it as if the coin were fair, then settle up. Risk-neutral pricing is that same diplomatic fiction for markets: to price a derivative, everyone agrees to compute as if the stock drifts at the boring risk-free rate, even though nobody believes that’s its true expected return.

Definition. Risk-neutral valuation says the fair price of a derivative today equals the expected payoff computed in a pretend world where every asset drifts at the risk-free rate $r$, discounted back at that same rate $r$. Formally:

where the expectation $\mathbb{E}^{\mathbb{Q}}$ is taken under the risk-neutral measure $\mathbb{Q}$ — the world where the drift is $r$, not the stock’s real expected return $\mu$.

Here’s the subtlety that trips everyone up, so let’s hit it head-on. When you simulated GBM paths in lesson 3, the drift you used was the asset’s actual expected return $\mu$. For pricing, you throw $\mu$ away and replace it with $r$. Why on earth would you simulate a world you don’t believe in?

Because of no-arbitrage. A derivative can be hedged by continuously trading the underlying and a risk-free bond. Once a payoff can be replicated by a self-financing hedge, its price is pinned down entirely by the cost of that hedge — and that cost has nothing to do with how bullish you feel about the stock. The hedging argument makes the real drift $\mu$ cancel out completely. What’s left is a world where every tradable grows at $r$ on average and you discount at $r$. Use the real $\mu$ instead and you’d be quoting a price nobody could hedge — free money for whoever takes the other side.

The Monte Carlo option-pricing recipe

Strip away the jargon and pricing any option by simulation is a four-line cookbook. Memorise the shape — every exotic in this lesson is the same four steps with a different payoff function in step 2.

1. Simulate $M$ risk-neutral paths. Roll $M$ GBM price paths from today’s spot to expiry $T$, using drift $= r$ (the risk-free rate) and the asset’s volatility $\sigma$. Each path is one full synthetic future — the whole journey, not just the endpoint.
2. Evaluate the payoff on each path. Apply the option’s payoff rule to every path. For a vanilla call that’s just $\max(S_T - K, 0)$; for an exotic it might read the path’s average, or check whether it ever crossed a barrier.
3. Average the payoffs. Take the plain arithmetic mean of those $M$ payoffs. That mean is your Monte Carlo estimate of the expected payoff under $\mathbb{Q}$.
4. Discount back to today. Multiply by $e^{-rT}$. That’s the price:

Beautifully, this works for a plain European option too — and when it does, the simulated price converges to the Black–Scholes formula as $M\to\infty$. That makes the vanilla a perfect sanity check: price it both ways, confirm they agree, and you’ve validated your simulator before you trust it on the exotic that has no formula to check against. (It’s also the basis of a control variate, a variance-reduction trick we’ll meet next lesson: lean on the vanilla’s known exact price to sharpen the exotic’s estimate.)

Watch the journeys: a fan of risk-neutral paths

Path-dependent payoffs need the whole thread, not just where it ends. The fan below is exactly the object step 1 produces — a batch of GBM futures, now drifting at the risk-free rate. Picture each thread two ways: an Asian option reads off the average height of the whole thread, and a barrier option asks whether any thread ever crossed a horizontal knock-out line. In both cases, the endpoint alone is silent — you have to keep the whole journey.

Asian options: paying on the average

Analogy. A vanilla option is judged by a single snapshot — the closing price on one specific day. An Asian option is judged by the whole semester’s grade average. One lucky (or rigged) final-exam print can’t make or break it, because the score is smeared across every observation in between.

Definition. An Asian option has a payoff based on the average price of the underlying over the life of the option, rather than the final price alone. For an Asian call with strike $K$, sampling the price at dates $S_1, S_2, \dots, S_n$:

Worked example. Suppose an Asian call is monitored at five dates and one simulated path prints:

The average is

With strike $K = 102$, the payoff on this path is $\max(104 - 102, 0) = 2$. Notice the final price was $104$, but a vanilla call would also have used a single point — the average smooths the spike at $112$ and the dip at $96$ into a single, calmer $104$. Repeat across $M$ paths, average those payoffs, multiply by $e^{-rT}$, and you have the Asian price.

Why it’s cheaper. Averaging throttles volatility. The average of several draws is less variable than any single draw (the classic “average of dice is tamer than one die”), so the effective volatility feeding the option is lower — and lower volatility means a cheaper option. There’s a business reason too: averaging makes the payoff hard to manipulate. A trader can sometimes nudge a single closing print to push a vanilla into the money; nudging a six-month average is far harder. That’s why Asians are popular in commodity and FX markets where a firm hedges a stream of purchases, not one dramatic settlement day.

Barrier options: the payoff with a trapdoor

Analogy. An up-and-out call is a sprinter who wins the race only if she never steps on a landmine. Cross the price barrier $H$ even once — for a single instant, mid-path — and the option is dead on the spot, no matter how gloriously it finishes. The final time on the clock is meaningless if she tripped the wire on the way.

Definition. A barrier option has a payoff that depends on whether the price path ever touched a barrier level $H$ during the option’s life. The two flavours:

• Knock-out — the option behaves like a vanilla unless the barrier is hit, in which case it pays 0 (it “knocks out”).
• Knock-in — the option pays nothing unless the barrier is hit, at which point it springs to life as a vanilla.

For an up-and-out call with strike $K$ and barrier $H$ (with $H > K$):

Worked example — two paths, same endpoint, different fates. Take strike $K = 100$, barrier $H = 120$, and two simulated paths:

Both paths end at $112$. A vanilla call would pay $12$ on each. But path A poked above $120$ on its way up, so the up-and-out call knocks out and pays nothing — while path B, which never breached the barrier, pays the full $12$. The endpoint is identical; the journeys are what split them. This is the single cleanest argument for full path simulation: you physically cannot tell A from B without watching every step.

Why barriers are cheaper than vanillas. A knock-out can only ever pay less than the vanilla (it sometimes dies early, never the reverse), so it’s strictly cheaper — which is precisely why clients buy them: the same upside exposure at a discount, in exchange for accepting the knock-out risk.

A quick cousin worth knowing: a lookback option pays on the path’s running maximum or minimum — e.g. a lookback call pays $\max(S_T - \min_t S_t,\,0)$, letting you “buy at the lowest price the stock ever printed.” Same lesson, different functional of the path: the endpoint alone can’t tell you the running min, so once again you simulate the whole journey.

Why this is the argument for simulating paths

Everything in this lesson points at one truth: for a path-dependent payoff, the journey carries information the endpoint physically does not. Store only $S_T$ and you’ve thrown away the average, the running max, and every barrier crossing — the very quantities the payoff is built from. There’s no reconstructing them after the fact; two paths with the same finish can have wildly different averages and opposite knock-out fates (you just saw it). Monte Carlo wins here not because it’s clever but because it’s the only method that keeps the whole path around to ask these questions of.

And these paths are nothing new — they’re the exact GBM fans from lesson 3, with one substitution: the drift is now $r$ instead of the real $\mu$, because we’re pricing, not forecasting. Same simulation engine, risk-neutral wiring.

Pitfalls: where simulated exotic prices go wrong

Simulation is powerful, but it has its own traps — and the nastiest one is specific to barriers.

Discretization bias on barriers. Monte Carlo checks the barrier only at the discrete time-steps it simulates. But the true path is continuous — it can dart above $H$ and back between two of your steps, a crossing your coarse grid never sees. The simulator then records “barrier not breached” when reality says it was, under-counting knock-outs and over-pricing the knock-out option. The fixes: simulate on a finer time grid (more steps = fewer missed crossings, at higher compute cost), or apply a Brownian-bridge correction — a formula for the probability the continuous path crossed $H$ between two simulated points, which patches the bias without brute-forcing the step count. Asians and lookbacks have milder versions of the same discrete-vs-continuous mismatch.

Garbage σ in, garbage price out. The universal Monte Carlo curse from the VaR lessons applies here in full force. Your option price is only as honest as the volatility you fed the simulator. A wrong $\sigma$ doesn’t yield a vague price — it yields a confidently precise wrong price, and barriers and Asians are more volatility-sensitive than vanillas (volatility governs how often paths kiss the barrier and how wild the average swings). Stress the input, not just the output.

No easy early exercise. Plain Monte Carlo prices European-style payoffs — exercise only at expiry — beautifully. It struggles with American options (exercisable any time before expiry) because at every step you’d need to compare “exercise now” against “the expected value of holding,” and that expected future value is itself unknown mid-simulation. The standard cure is Longstaff–Schwartz least-squares Monte Carlo, which regresses the continuation value from the simulated paths — powerful, but extra machinery beyond this lesson.

Putting it together

To price an option by simulation, you compute in a pretend risk-neutral world: simulate $M$ GBM paths with drift $= r$ (not the real $\mu$ — no-arbitrage makes the real drift cancel), evaluate the payoff on each path, average them, and discount at $e^{-rT}$. For a vanilla this just rediscovers Black–Scholes — a clean sanity check. But the method earns its keep on path-dependent payoffs no formula can touch: Asian options that read the path’s average (cheaper, harder to rig), and barrier options that live or die on whether the path ever touched a level. Two paths can share an endpoint and have opposite fates, which is the whole reason you must simulate the journey, not just the destination. Mind the traps — discretization bias on barriers, garbage-$\sigma$-in, and no easy American exercise — and you can value almost any exotic on the menu.

Next up — Convergence and Variance Reduction — we get quantitative about the price tag of accuracy: how many paths you actually need before a simulated price is trustworthy, why the error only shrinks like $1/\sqrt{M}$, and the clever tricks (antithetic variates, control variates, and the vanilla-as-control-variate idea we teased here) that buy the same precision with a fraction of the paths.