Monte Carlo Estimation for Asian Options

Why Asian Options Require Path Simulation

A European call depends on the asset price at one date. An arithmetic-average Asian call depends on prices at several dates, so each simulated outcome must retain the relevant path values rather than only the terminal value. The averaging convention matters: including or excluding the initial price, using equally spaced dates or irregular dates, and using a discrete or continuous average define different payoffs.

Work on a filtered probability space $(\Omega,\mathcal{F},(\mathcal{F}_t)_{t\ge0},\mathbb{P})$ carrying a Brownian motion $B_t$ under $\mathbb{P}$. For pricing, suppose an equivalent risk-neutral measure $\mathbb{Q}$ is specified, under which $B_t^{\mathbb{Q}}$ is Brownian motion. Expectations for prices are taken under $\mathbb{Q}$, not $\mathbb{P}$. For a non-dividend-paying asset with constant interest rate $r$, use the risk-neutral model $dS_t=rS_t\,dt+\sigma S_t\,dB_t^{\mathbb{Q}}$, with $S_0>0$ and $\sigma\ge0$. In stochastic-integral notation, integrands are adapted, progressively measurable, and belong to $\mathcal{L}^2$ on the time interval in question; under $\mathbb{Q}$, this means $\mathbb{E}^{\mathbb{Q}}\int_0^T H_s^2\,ds<\infty$. The drift is $rS_t$, and the diffusion coefficient is $\sigma S_t$.

Definition: Fix $T>0$, an integer $m\ge1$, and equally spaced monitoring dates $t_j=jT/m$, $j=1,\ldots,m$. The discretely monitored arithmetic average and discounted Asian call payoff are $$ A_m=\frac{1}{m}\sum_{j=1}^{m}S_{t_j}, \qquad Y_m=e^{-rT}(A_m-K)^+, $$ where $K\ge0$ is the strike and $x^+=\max(x,0)$. The risk-neutral price is $V_m=\mathbb{E}^{\mathbb{Q}}[Y_m]$, when this expectation is finite. This convention excludes $S_0$ from the average.

For a simulated path, first record all $m$ monitored prices, average them, subtract the strike, take the positive part, and discount once from maturity to time zero. The discounting is by $e^{-rT}$, not by a separate factor at each monitoring date: the payoff is received at $T$.

The Monte Carlo Estimator and Its Justification

Generate $N$ independent paths under $\mathbb{Q}$, and let $Y_m^{(i)}$ be the discounted payoff from path $i$. The Monte Carlo estimator is the sample mean of these payoffs. It estimates the discretely monitored price $V_m$; it does not automatically estimate the price of a continuously averaged option or remove error from approximating the diffusion paths numerically.

Theorem: If the simulated payoffs $Y_m^{(1)},\ldots,Y_m^{(N)}$ are independent and identically distributed as $Y_m$, with $\mathbb{E}^{\mathbb{Q}}[Y_m^2]<\infty$, then $$ \widehat V_{m,N}=\frac1N\sum_{i=1}^NY_m^{(i)},\qquad \mathbb{E}^{\mathbb{Q}}[\widehat V_{m,N}]=V_m,\qquad \operatorname{Var}^{\mathbb{Q}}(\widehat V_{m,N})=\frac{\operatorname{Var}^{\mathbb{Q}}(Y_m)}{N}. $$

Proof. Each payoff has finite second moment and therefore finite first moment. Linearity of expectation gives $\mathbb{E}^{\mathbb{Q}}[\widehat V_{m,N}]=N^{-1}\sum_i\mathbb{E}^{\mathbb{Q}}[Y_m^{(i)}]=V_m$. For the variance, independence makes each covariance between distinct payoffs zero, so $\operatorname{Var}^{\mathbb{Q}}(\widehat V_{m,N})=N^{-2}\sum_i\operatorname{Var}^{\mathbb{Q}}(Y_m^{(i)})=\operatorname{Var}^{\mathbb{Q}}(Y_m)/N$. Thus the estimator is unbiased, and its standard deviation is the payoff standard deviation divided by $\sqrt{N}$. $\square$

In a run with $N\ge2$, the payoff variance is unknown and is estimated by $s_Y^2=(N-1)^{-1}\sum_i(Y_m^{(i)}-\widehat V_{m,N})^2$. The estimated standard error of the sample mean is $s_Y/\sqrt{N}$. This quantifies sampling uncertainty for the chosen discretely monitored payoff. Any bias from choosing a finite number of monitoring dates, or from a numerical time-stepping scheme, is a separate issue.

Worked Example: Payoffs and Standard Error from Three Paths

Take $T=1$, $r=0$, $K=100$, and three monitoring dates. Suppose three simulated paths produce the monitored prices $(100,110,120)$, $(100,90,80)$, and $(100,105,105)$, respectively. Compute the estimator and its estimated standard error.

The three averages are $(100+110+120)/3=330/3=110$, $(100+90+80)/3=270/3=90$, and $(100+105+105)/3=310/3$. Since the discount factor is $e^{-0\cdot1}=1$, the discounted payoffs are $10$, $0$, and $10/3$. Therefore $$ \widehat V_{3,3}=\frac{10+0+10/3}{3}=\frac{40/3}{3}=\frac{40}{9}. $$

For the sample variance, the deviations from $40/9$ are $50/9$, $-40/9$, and $-10/9$. Their squared sum is $(2500+1600+100)/81=4200/81=1400/27$. Dividing by $N-1=2$ gives $s_Y^2=700/27$, so the estimated standard error is $\sqrt{(700/27)/3}=\sqrt{700/81}=10\sqrt7/9$.

Check: the squared deviations sum to $1400/27$, and the sample variance $700/27$ times $N-1=2$ recovers that sum. The sample mean is $40/9$, between the smallest payoff $0$ and largest payoff $10$, as it must be.

Simulating the Monitored Prices

For geometric Brownian motion, exact transitions can be used at the monitoring dates rather than taking smaller Euler steps. The solution over a step of length $h$ is $S_{t+h}=S_t\exp((r-\sigma^2/2)h+\sigma(B_{t+h}^{\mathbb{Q}}-B_t^{\mathbb{Q}}))$. To verify it, apply Ito's formula to $f(t,b)=S_0\exp((r-\sigma^2/2)t+\sigma b)$. Its derivatives are $f_t=(r-\sigma^2/2)f$, $f_b=\sigma f$, and $f_{bb}=\sigma^2f$. The Ito correction is $\tfrac12f_{bb}\,dt=\tfrac12\sigma^2f\,dt$, so the drift becomes $rf\,dt$ and the diffusion term is $\sigma f\,dB_t^{\mathbb{Q}}$, as required. This verification uses $dB_t\,dB_t=dt$, $dB_t\,dt=0$, and $dt\,dt=0$.

At each step, the Brownian increment is generated as $\Delta B_j^{\mathbb{Q}}=\sqrt{h}\,Z_j$, where the $Z_j$ are independent standard normal variables. The same simulated path must be used to produce all monitoring prices in its average.

Worked Example: Exact Geometric Brownian Updates

Let $S_0=100$, $r=1/2$, $\sigma=1$, $T=2$, and $m=2$, so $h=1$. Consider a simulated path whose two Brownian increments are $\log 2$ and $-\log 2$. Compute its Asian call payoff for $K=125$.

Since $r-\sigma^2/2=1/2-1/2=0$, the first update is $S_1=100e^{\log2}=200$. The second is $S_2=200e^{-\log2}=100$. Thus the average at the two monitoring dates is $(200+100)/2=150$, the undiscounted payoff is $(150-125)^+=25$, and the discounted payoff is $e^{-(1/2)2}\cdot25=25/e$.

Check: the product of the two price multipliers is $e^{\log2-\log2}=1$, so the terminal price returns to $100$. The average still equals $150$, confirming why the Asian payoff cannot be computed from the terminal price alone. The inputs are specified Brownian increments for this calculation; the formula does not assert that a continuously distributed increment equals either specified value with positive probability.

Increasing Monitoring Frequency

A continuously averaged arithmetic payoff uses $A=T^{-1}\int_0^T S_t\,dt$ instead of a finite sum. For a continuous price path, right-endpoint averages $A_m$ converge to this integral average by Riemann-sum convergence. Because $x\mapsto(x-K)^+$ is continuous and satisfies $|(x-K)^+-(y-K)^+|\le|x-y|$, the corresponding pathwise payoffs converge too.

Proposition: Let $S$ have continuous paths on $[0,T]$, let $K\ge0$, and suppose $\mathbb{E}^{\mathbb{Q}}[(\sup_{0\le t\le T}S_t)^+]<\infty$. Define $A=T^{-1}\int_0^T S_t\,dt$, $Y=e^{-rT}(A-K)^+$, and $Y_m=e^{-rT}(A_m-K)^+$ using the equally spaced right-endpoint average. Then $Y_m\to Y$ almost surely and $\mathbb{E}^{\mathbb{Q}}[Y_m]\to\mathbb{E}^{\mathbb{Q}}[Y]$.

Proof. For every continuous path, its right-endpoint Riemann sums $A_m$ converge to $A$. Continuity of the positive-part function therefore gives $Y_m\to Y$ pathwise. To pass to expectations, observe that $A_m\le\sup_{0\le t\le T}S_t$ and $A_m-K\le A_m$, since $K\ge0$. Hence $0\le Y_m\le e^{-rT}(\sup_{0\le t\le T}S_t)^+$. This is an integrable bound by assumption. The same bound applies to $Y$. The dominated convergence theorem now gives convergence of expectations. The positive part on the supremum is necessary: a continuous path need not be nonnegative under the proposition's assumptions. $\square$

This proposition separates two effects. Increasing $m$ reduces the difference between the discrete and continuous monitoring conventions under its assumptions. It does not reduce Monte Carlo sampling error by itself, and it does not guarantee monotone prices as $m$ changes. If paths are generated with a numerical SDE scheme, the error from that scheme also remains distinct.

Worked Example: Monitoring Frequency on a Linear Path

On $[0,1]$, let $S_t=80+40t$, with strike $K=100$. Compare right-endpoint averages for $m=2$ and $m=4$ with the continuous average. Take $r=0$, so discounting does not change the payoffs.

For $m=2$, the monitoring values are $S_{1/2}=80+40(1/2)=100$ and $S_1=80+40=120$. Thus $A_2=(100+120)/2=110$, and the payoff is $(110-100)^+=10$. For $m=4$, the values are $90,100,110,120$, obtained by substituting $t=1/4,1/2,3/4,1$. Their average is $420/4=105$, giving payoff $5$.

The continuous average is $\int_0^1(80+40t)\,dt=80+20=100$, so its payoff is $0$. More generally, the right-endpoint average is $$ A_m=\frac1m\sum_{j=1}^m\left(80+\frac{40j}{m}\right) =80+\frac{40}{m^2}\frac{m(m+1)}2 =100+\frac{20}{m}. $$

As $m\to\infty$, this tends to $100$, and the corresponding payoff $(20/m)^+$ tends to $0$. Check: substituting $m=2$ gives $100+20/2=110$, and substituting $m=4$ gives $100+20/4=105$, matching the direct averages.

Common Pitfalls

Using only the terminal price. For an Asian call, the average is computed from every specified monitoring value. In the geometric Brownian example, $S_2=100$, which alone would give a zero call payoff at strike $125$; the two-date average is $150$, giving the positive payoff $25$ before discounting.

Changing the averaging convention silently. The definition here excludes $S_0$ and includes $S_T$. If the initial price is included or monitoring dates are irregular, the weights in the average must be changed and the estimator must use that stated convention.

Confusing sampling error with monitoring error. The standard error estimates variability from using finitely many simulated paths for a fixed payoff definition. The difference between a discrete average and a continuous average is a separate approximation effect. More simulated paths address the former, while increasing monitoring frequency addresses the latter under appropriate path regularity.

Check Your Understanding

Use the stated convention that monitoring dates exclude the initial price and include maturity.

For monitored prices $80,100,120$, strike $95$, and $r=0$, compute the three-date arithmetic-average call payoff.
State the finite-second-moment condition needed for the Monte Carlo variance formula, and explain why independence of paths matters.
For geometric Brownian motion, write the exact one-step update over a monitoring interval of length $h$ and identify the Brownian increment's distribution under $\mathbb{Q}$.
Explain why convergence of the discrete average on each continuous path does not alone imply convergence of expected payoffs.
For $S_t=80+40t$ on $[0,1]$, compute the right-endpoint average when $m=5$, and compare it with the continuous average.