2.7 ended with a striking result: the option price satisfies a PDE that contains no trace of the stock's real-world drift μ\mu. But a Monte Carlo pricer still needs a drift in its simulator, and the obvious choice, the stock's expected return, gives the wrong answer. With μ=0.1\mu = 0.1 the discounted average payoff is 21.42, not the option's price of 18.02; at higher drifts the gap becomes larger still.

This page answers the question that creates. If the real-world drift does not belong in the price, what drift does? The answer is rr. Simulate the stock as if it drifted at the risk-free rate, average the payoff and discount, and the price appears. That sounds wrong at first glance: the stock does not drift at rr, and nobody believes it does.

The point of the page is to show why this apparent mistake is the correct calculation. The reason is the hedge from 2.7. Once the option is combined with the right position in the stock, the resulting portfolio earns rr whatever the stock's expected return may be. The hedge removes μ\mu from the problem, and with it removes any need to know what investors think the stock will earn.

That change in viewpoint leads to the central formula of the page. The price is not the discounted expectation under the stock's real-world probabilities. Instead it is the discounted expectation in a different measure, one in which the discounted stock is a martingale. The page shows that this is not a different stock or a different set of paths, but the same uncertainty viewed through different weights. The hedge licenses the switch, Girsanov explains it, and the result is the pricing recipe used in practice: simulate under rr, report a standard error, and treat the price as a risk-neutral expectation.

What this page shows, and how

This page answers three practical questions that appear as soon as pricing is written as a simulation.

Which drift belongs in the simulator? The obvious answer is the stock's expected return: simulate the stock at μ\mu, average the payoff, discount at rr. The results show that this is wrong. At μ=0.1\mu = 0.1 the discounted average payoff is 21.42 against a price of 18.02, and the gap becomes larger as μ\mu moves further from rr. The other tempting answer, that any drift will do because μ\mu disappeared from the PDE, is wrong as well. This section shows that only the rr-drift produces the price, and explains why the hedge of 2.7 licenses that choice.

How many paths are enough, and how does a Monte Carlo pricer know whether its answer is trustworthy? The wrong instinct is to treat the Monte Carlo estimate as the price itself, or to imagine that ten times as many paths give ten times the precision. Neither is true. A Monte Carlo run returns a price and a standard error, the uncertainty shrinking only like 1/n1/\sqrt{n}. This section measures that law directly and checks that the standard error reported by the pricer is the error it actually makes.

Must paths generated under the wrong drift be discarded? A natural answer is yes: if pricing requires the rr-drift, surely paths drawn at μ\mu have to be regenerated. Another common misconception is that changing the drift changes the volatility as well. This section shows that neither is true. The same paths can be reused by reweighting them, the volatility remains unchanged, and the weights move the average from the real-world measure to the risk-neutral one. That measured reweighting is Girsanov's theorem in practice.

The experiment — what was done

The experiment follows the decisions already made in 2.6 and 2.7. Because a European payoff depends only on the terminal stock price, the stock is never stepped by Euler and never approximated: each terminal value is drawn exactly from the closed-form solution of 2.6's Geometric Brownian motion (GBM), ST=S0⋅e(μ−12σ2)T+σWTS_T = S_0 \cdot e^{(\mu - \frac12\sigma^2)T + \sigma W_T}. The main drift test uses 10610^6 independent draws at each of six values of μ\mu: −0.2, 0, 0.05, 0.1, 0.3 and 0.6. For each run the discounted average payoff is compared with the option's price, and the forward is carried alongside as a second product whose value is already known from no-arbitrage.

Where paths are needed rather than terminal values, the stock is simulated with 2.6's exact step. The experiment generates 10510^5 paths at three drifts, μ=r\mu = r, μ=0.3\mu = 0.3 and μ=−0.2\mu = -0.2, and uses them to examine the discounted stock through time, the martingale property, and Girsanov reweighting. The same paths are viewed both as drawn and after reweighting, so any change can be attributed to the weights rather than to a new simulation.

A benchmark price is needed throughout. This page uses the Black–Scholes closed form as the solution of 2.7's PDE with the payoff as its condition at TT, because 2.7 already checked directly that the formula satisfies the PDE. That formula will be derived in 2.9; here it is borrowed as a known solution so that the Monte Carlo results can be compared against a fixed target.

Finally, the page builds a Monte Carlo surface VMC(t,S)V_{\text{MC}}(t, S) over time and stock price. For each point (t,S)(t, S), the risk-neutral expectation is estimated and compared with the closed form. This lets the experiment test not only whether a single price is correct, but whether the expectation and the PDE describe the same function across a grid of (t,S)(t, S).

·

Which drift belongs in the simulator?

The first question is the simplest one to ask and the easiest one to answer incorrectly: what drift belongs in the simulator? The natural choice is the stock's expected return. Simulate the stock at μ\mu, average the payoff, discount at rr, and call the result the price. The table shows that this is not what pricing produces. At μ=0.1\mu = 0.1 the discounted average payoff is 21.42 against a price of 18.02; at μ=0.3\mu = 0.3 it rises to 38.96; at μ=−0.2\mu = -0.2 it falls to 6.77. The discrepancy is not Monte Carlo noise. Every row matches its own closed-form expectation and every measured mean stock price matches S0eμTS_0 e^{\mu T}. The simulator is doing exactly what it was asked to do. The problem is that the discounted real-world expectation is not the price.

Discounted payoff at six drifts

10610^6 exact draws at each drift, S0=K=100S_0 = K = 100, r=0.05r = 0.05, σ=0.4\sigma = 0.4, T=1T = 1. Seed 20261104, from the committed experiments/2_8_risk_neutral_pricing.json: results.terminal.mu_*.{discounted_mean_payoff,closed_form_discounted_real_world_expectation,measured_bias,bias_in_se,E_S_T,forward_discounted_mean,forward_closed_form_under_mu}. The price is results.closed_form_price, 18.023. The figures in brackets and the closed-form column are formulas, not measurements.

drift μ\mue−rT⋅e^{-rT} \cdot mean payoff ± s.e.closed form under μ\mugap to the price 18.02in s.e.E[ST]\E[S_T] (S0eμTS_0 e^{\mu T})forward: e−rT⋅e^{-rT} \cdot mean(ST−K)(S_T - K) (closed form)
−0.26.77 ± 0.026.74−11.25−62281.92 (81.87)−17.20 ± 0.03 (−17.24)
015.11 ± 0.0315.08−2.92−103100.06 (100.00)0.06 ± 0.04 (0.00)
0.05 = rr18.01 ± 0.0318.02−0.01−0.4105.11 (105.13)4.86 ± 0.04 (4.88)
0.121.42 ± 0.0321.36+3.40+99110.58 (110.52)10.06 ± 0.04 (10.00)
0.338.96 ± 0.0539.01+20.93+439134.91 (134.99)33.21 ± 0.05 (33.28)
0.679.64 ± 0.0779.69+61.62+877182.18 (182.21)78.17 ± 0.07 (78.20)

Changing only the drift to rr makes the same code return 18.01 ± 0.03, within sampling error of the benchmark value 18.02. Nothing else changes: the payoff, volatility, discounting and random draws are all treated in the same way. The lesson is not that the stock somehow drifts at rr. It is that the hedge of 2.7 removes μ\mu from the pricing problem, leaving one special drift under which a discounted expectation reproduces the value implied by no-arbitrage.

The forward makes the same point more sharply because its value is already known. No simulation is needed to show that the forward is worth S0−Ke−rT=4.88S_0 - Ke^{-rT} = 4.88. Yet the discounted expected payoff of the forward depends on μ\mu and agrees with 4.88 only when the simulation is run at the rr-drift. The risk-neutral expectation therefore is not a special trick for options. It is the expectation that reproduces the no-arbitrage value even in the simplest derivative where the answer is known in advance.

The distinction between expectation and price appears most clearly when different desks disagree about μ\mu. A desk using μ=0.3\mu = 0.3 expects the call to pay 40.95 on average; a desk using μ=−0.2\mu = -0.2 expects only 7.12. They disagree almost completely about what the option will pay. Yet they still arrive at the same price. The expected payoff depends on what a desk believes about the stock. The price depends on what it costs to hedge, and 2.7 showed that the hedge does not depend on μ\mu at all.

How many paths are enough, and how does a Monte Carlo pricer know whether its answer is trustworthy?

The second question is not whether a Monte Carlo pricer can produce the right answer, but how accurately it has done so. A Monte Carlo estimate is always a price with an error bar attached. As the number of paths increases from 10310^3 to 10610^6, the reported uncertainty falls from ±0.95 to ±0.31, ±0.10 and finally ±0.03. The improvement is steady but slow: each hundred-fold increase in paths buys roughly one extra decimal place of accuracy. The wrong expectation is that more paths improve precision linearly. They do not. Averaging independent draws reduces uncertainty only like 1/n1/\sqrt{n}.

The experiment measures that law directly. The 10610^6-path risk-neutral run is cut into smaller independent batches and treated as many separate pricers. One hundred pricers of 10410^4 paths each produce estimates whose rms error around the benchmark price is 0.298. Each pricer reports a standard error of 0.312. At 10210^2 paths the measured scatter is 3.16 and the reported value is 3.12. The quantity printed by the pricer is therefore not a guess about its accuracy but a measured prediction of the error made by runs of that size.

Error against n, nested and batched

The 10610^6 draws at μ=r\mu = r, read two ways, from results.terminal.mu_0.05.convergence: the first nn of them as one estimate, nested[*].{n,estimate,se,error_in_se}, and all of them cut into independent batches of nn, batches[*].{n,batches,rms_error,predicted_se,forward_rms_error,forward_predicted_se}. The reported s.e. in the batch columns is sd⁡/n\sd/\sqrt{n} with the sd⁡\sd of the full run. It is the s.e. a pricer of nn paths reports. Batches stop at 10,000 so that each row has at least 100 of them.

nn pathsrisk-neutral estimate ± s.e. (nested)error, in s.e.independent batches of nnrms error of the batch meanss.e. a pricer reports, sd⁡/n\sd/\sqrt{n}forward: rms error (reported s.e.)
100——10,0003.163.124.19 (4.16)
1,00018.22 ± 0.95+0.21,0000.9750.9881.29 (1.32)
10,00018.35 ± 0.31+1.11000.2980.3120.385 (0.416)
100,00018.06 ± 0.10+0.4————
1,000,00018.010 ± 0.031−0.4————

The same pattern appears for the forward, whose payoff contains none of the option's non-linearity. The rms error falls with a log-log slope of about −0.51 for the call and −0.52 for the forward, matching the expected −12-\tfrac12 law. The square-root behaviour comes from averaging, not from any special property of options.

The practical conclusion is simple. A Monte Carlo calculation never returns a price on its own. It returns a price and a standard error. Quoting one without the other is incomplete, because the standard error is the information that tells the reader how much confidence to place in the estimate. The point of the simulation is not to eliminate uncertainty but to measure it.

Must paths generated under the wrong drift be discarded?

The third question is whether paths simulated under the wrong drift must be thrown away. The answer is no. If paths were drawn under a real-world drift μ\mu, each path is assigned the weight

L=e−λWT−12λ2T,λ=(μ−r)/σ.L = e^{-\lambda W_T - \frac12\lambda^2 T}, \quad \lambda = (\mu - r)/\sigma.

This is the same exponential martingale that appeared in 2.5, now used to redistribute probability across the existing paths. Its mean is 1, so it does not add or remove mass from the average. Instead it changes which paths matter most.

The effect is visible immediately in the prices. The paths drawn at μ=0.3\mu = 0.3 produce a discounted average payoff of 38.96 on their own; with the weights applied, the same paths produce 18.00 ± 0.01. The μ=−0.2\mu = -0.2 paths move from 6.77 to 18.11 ± 0.07. No path has changed and no new simulation has been run. The average changes because the weights shift probability away from outcomes that are overrepresented under μ\mu and towards outcomes that are underrepresented.

Terminal draws reweighted by L

The same 10610^6 draws as the first table, each weighted by LL, from results.terminal.mu_*: lambda, discounted_mean_payoff.mean, itm_fraction_P.mean and girsanov.{reweighted_discounted_mean_payoff,reweighted_forward,mean_L,ess_over_n,expected_ess_over_n,reweighted_itm_fraction}. The bracketed e−λ2Te^{-\lambda^2 T} is the formula, not a measurement. The risk-neutral in-the-money probability is itm_probability_Q, 0.4701. The in-the-money column gives the share of paths, then the share of weight.

paths drawn at μ\muλ\lambdaas drawn: e−rT⋅e^{-rT} \cdot mean payoffreweighted by LL ± s.e.reweighted forward (4.88)mean LLeffective sample size / nn (e−λ2Te^{-\lambda^2 T})in the money: paths → weight
−0.2−0.6256.7718.11 ± 0.074.98 ± 0.081.001 ± 0.0010.676 (0.677)24.3 % → 47.2 %
0−0.12515.1118.05 ± 0.044.93 ± 0.041.000 ± 0.0000.985 (0.984)42.2 % → 47.1 %
0.1+0.12521.4218.07 ± 0.034.92 ± 0.041.000 ± 0.0000.984 (0.984)52.0 % → 47.0 %
0.3+0.62538.9618.00 ± 0.014.81 ± 0.041.001 ± 0.0010.677 (0.677)70.8 % → 47.0 %
0.6+1.37579.6418.05 ± 0.015.25 ± 0.110.993 ± 0.0020.158 (0.151)90.4 % → 47.1 %

The reweighting can be seen directly in the distribution. Under μ=0.3\mu = 0.3, 70.8 % of paths finish in the money. After reweighting, 47.0 % of the probability mass finishes in the money, matching the risk-neutral probability of finishing in the money, 47.0 %. The μ=−0.2\mu = -0.2 run moves in the opposite direction, from 24.3 % to 47.2 %. The weights transform the μ\mu-sample into a QQ-sample without drawing a single new random number.

Girsanov’s weights in the distribution. ST drawn at μ = 0.3 and at μ = −0.2, as drawn and reweighted by L, against a run drawn at r.
drawn at μ = 0.30K = 100200300400
drawn at μ = −0.20K = 100200300400
  • as drawn at μ
  • reweighted by L
  • drawn at r, dashed
Each line is a run’s share of its draws, or of its weight, in each band of ST 5 wide: 106 draws a run, in 80 bands from 0 to 400. Right of the dashed line at K is in the money. Reweighted, each run is within 0.09 percentage points of the run drawn at r in every band.

Just as importantly, the weights change the drift and not the volatility. For the μ=0.3\mu = 0.3 paths, the mean log-return moves from 0.222 to −0.028, against (r−12σ2)T=−0.030(r - \tfrac12\sigma^2)T = -0.030. The realised volatility moves only from 0.3999 to 0.3994. The paths are the same, the noise is the same and σ\sigma is the same. What changes is the probability assigned to each path. Girsanov changes the measure, not the volatility.

Log-return and volatility, reweighted

10510^5 paths of 256 exact steps at each of three drifts, the weight LL taken at TT, from results.paths.mu_*: log_return_T.{unweighted,expected_under_mu,reweighted,expected_under_Q} and realised_variance.{realised_vol_unweighted,realised_vol_reweighted}. At μ=r\mu = r the weight is 1, so there is nothing to reweight. The figures in brackets are the formula, not a measurement.

stepped paths (105×25610^5 \times 256) drawn at μ\mumean log⁡(ST/S0)\log(S_T/S_0): as drawn ((μ−12σ2)T(\mu - \tfrac12\sigma^2)T)reweighted ((r−12σ2)T=−0.030(r - \tfrac12\sigma^2)T = -0.030)realised volatility: as drawnreweighted
0.05 = rr−0.030 ± 0.001 (−0.030)—0.4001—
0.3+0.222 ± 0.001 (+0.220)−0.028 ± 0.0020.39990.3994
−0.2−0.281 ± 0.001 (−0.280)−0.030 ± 0.0020.40000.3998

The price paid for this convenience is statistical rather than theoretical. The weights make some paths count more than others, reducing the effective sample size from nn to approximately

n⋅e−λ2T.n \cdot e^{-\lambda^2 T}.

When μ\mu is close to rr almost every path still counts in full; when μ\mu is far from rr, many paths contribute little weight and the effective number of paths falls. The answer remains correct, but fewer effective observations are doing the work.

A practical consequence, and no more than an aside here, is that some choices of μ\mu can accidentally reduce the variance of a particular estimator. For this call option the μ=0.3\mu = 0.3 paths produce a smaller standard error after reweighting than the direct QQ simulation. This is the basic idea behind importance sampling, which 3.2 takes up with the other variance-reduction methods. The central point on this page is simpler: changing from μ\mu to rr does not require new paths, and it does not change σ\sigma. It changes the weight attached to each path, and those weights are enough to recover the risk-neutral price.

Why — from the hedge to the expectation

The numbers in the previous sections all point to the same conclusion; this section explains why it is true. 2.7 ended with a hedge whose return is rr whatever value μ\mu takes. Once the option is combined with the correct holding in the stock, the drift of the stock disappears from the hedged book, and the resulting PDE contains no μ\mu. That means the option's value cannot depend on the market's opinion of the stock's expected return. The stock still has a real-world drift, but the price does not.

If the price is independent of μ\mu, the natural next step is to choose whichever drift makes pricing easiest. The PDE from 2.7 already contains rr where a drift term would appear, so we value the option as if the stock followed

dS=rS dt+σS dW~.\dS = rS \,\dt + \sigma S \,\dd\tilde{W}.

This is not a claim about the stock's behaviour. It is a change of viewpoint. The stock in the market still drifts at μ\mu; pricing is performed under a measure in which the drift is set to rr.

Under that choice, the discounted stock becomes especially simple. Substituting rr into 2.6's solution gives

e−rtSt=S0⋅eσW~t−12σ2t,e^{-rt}S_t = S_0 \cdot e^{\sigma\tilde{W}_t - \frac12\sigma^2 t},

which is exactly the exponential martingale from 2.5. Its expectation remains S0S_0 at every time. Measured on the paths, that is what a martingale must show in the checks below: the discounted stock stays centred on 100 while its spread grows.

The next step is to write the option price itself as an expectation. Define

V(t,S)=e−rτ⋅EQ[payoff∣St=S].V(t, S) = e^{-r\tau} \cdot \E^Q[\text{payoff} \mid S_t = S].

As time passes, the discounted conditional expectation of a fixed future payoff is a martingale. A martingale has no drift term. Applying Itô's lemma to e−rtV(t,St)e^{-rt}V(t, S_t) under the rr-drift and setting the drift to zero produces exactly the PDE from 2.7. This is the content of the Feynman–Kac theorem: the risk-neutral expectation and the PDE are two descriptions of the same function. The Monte Carlo surface in the checks below checks that relationship directly.

The remaining question is how the QQ-measure is related to the real-world one. Girsanov's theorem answers that. If

λ=(μ−r)/σ,\lambda = (\mu - r)/\sigma,

then the weight

L=e−λWT−12λ2TL = e^{-\lambda W_T - \frac12\lambda^2 T}

moves expectations from PP to QQ:

EQ[X]=EP[L⋅X].\E^Q[X] = \E^P[L \cdot X].

Nothing about the paths themselves changes. The same simulated paths are reused, volatility remains σ\sigma, and only the weight attached to each path changes. The reweighting section measured exactly that effect: the weighted drift moves from the real-world value to the risk-neutral one while the realised volatility is unchanged to one part in a thousand.

This is why the measure is called risk-neutral. Under QQ, discounted traded assets have no drift and therefore behave like martingales. The name describes the property of the measure, not the reason it exists. The reason is the hedge from 2.7. Risk-neutral valuation is not a statement about investors, psychology or beliefs about the stock. It is the expectation formula forced by no-arbitrage once the hedge has removed μ\mu from the pricing problem.

The martingale and the PDE checked

The final checks return to the two claims that support the whole construction. First, under the rr-drift the discounted stock should be a martingale. Second, the risk-neutral expectation should be the same function as the solution of 2.7's PDE with the payoff as its condition at TT. Both can be measured directly.

The martingale check is the simpler of the two. Under QQ, the mean of the discounted stock stays at 100 throughout the simulation: 99.98, 100.00, 99.99 and 100.01 at the quarter-year checkpoints. The paths spread out over time, but their discounted mean does not move. Under the real-world drift they do. At μ=0.3\mu = 0.3 the discounted stock climbs to 128.6 by maturity; at μ=−0.2\mu = -0.2 it falls to 77.8. Those are exactly the behaviours expected when the discounted stock drifts at μ−r\mu - r rather than zero. Applying the Girsanov weights restores the martingale on the same paths: the μ=0.3\mu = 0.3 and μ=−0.2\mu = -0.2 runs return values close to 100 at every checkpoint and 99.9 by maturity. The martingale is therefore a property of the measure, not of the paths themselves.

Discounted stock at each checkpoint

The same 10510^5 stepped paths, from results.paths.mu_*.checkpoints[*]: discounted_S.{mean,se_of_mean,sd}, expected_under_mu, expected_sd_under_mu and reweighted_discounted_S.{mean,se_of_mean}. The figures in brackets are the formulas, not measurements; in the μ\mu rows the bracketed sd⁡\sd carries the factor e(μ−r)te^{(\mu - r)t} as well.

E[e−rtSt]\E[e^{-rt}S_t] ± s.e.t=14t = \tfrac14t=12t = \tfrac12t=34t = \tfrac34t=1t = 1sd⁡\sd at t=1t = 1 (S0eσ2t−1S_0\sqrt{e^{\sigma^2 t} - 1})
μ=r\mu = r (QQ)99.98 ± 0.06100.00 ± 0.0999.99 ± 0.11100.01 ± 0.1341.8 (41.7)
μ=0.3\mu = 0.3, expected S0e(μ−r)tS_0 e^{(\mu - r)t}106.50 ± 0.07 (106.45)113.45 ± 0.10 (113.31)120.85 ± 0.14 (120.62)128.65 ± 0.17 (128.40)53.6 (53.5)
μ=−0.2\mu = -0.293.89 ± 0.06 (93.94)88.14 ± 0.08 (88.25)82.86 ± 0.09 (82.90)77.81 ± 0.10 (77.88)32.5 (32.4)
μ=0.3\mu = 0.3, reweighted by LL99.75 ± 0.1999.81 ± 0.1699.85 ± 0.1299.90 ± 0.07—
μ=−0.2\mu = -0.2, reweighted by LL99.83 ± 0.2799.87 ± 0.3399.98 ± 0.3899.85 ± 0.43—

The second check evaluates the pricing function itself. A Monte Carlo surface VMC(t,S)V_{\text{MC}}(t, S) is constructed on sixteen stock prices and two times to expiry, τ=1\tau = 1 and 12\tfrac12, giving thirty-two points in (t,S)(t, S)-space. At each point the option value is estimated from the risk-neutral expectation and compared with the closed form. The largest disagreement is only 1.7 standard errors, with most points appreciably closer. Across the whole grid, the expectation and the closed form are statistically indistinguishable.

A stronger test is obtained by treating the Monte Carlo surface as if it were the true pricing function and substituting finite-difference estimates of its derivatives into 2.7's PDE. The resulting residual is not exactly zero, but it is the same size as the Monte Carlo noise measured independently by replication. In other words, the discrepancy is no larger than the noise expected from the simulation itself. The expectation therefore satisfies the PDE to Monte Carlo precision, while the closed form satisfies it to numerical accuracy.

Monte Carlo surface against the PDE

VMCV_{\text{MC}} against the closed form on the grid, h=2h = 2 % of SS and of τ\tau, from results.feynman_kac.rows[*]: max_abs_error_in_se, pde_residual_rms, pde_residual_max_abs, closed_form_fd_residual_rms, mc_noise_rms_by_replication and largest_pde_term_rms. The worst-error and residual columns read the first of two replications, and the noise column is the difference between the two. The closed form's analytic residual on the same grid, analytic_residual_max_abs, is under 3×10−153 \times 10^{-15}. Each row is 16 values of SS from 50 to 200, 10610^6 common draws each. The closed form's own residual is its finite-difference residual, at the same h=2h = 2 %. The Monte Carlo noise is in the residual, by replication.

Monte Carlo surface VMC(t,S)V_{\text{MC}}(t, S)worst ∣VMC−V∣\lvert V_{\text{MC}} - V \rvert in s.e.PDE residual on VMCV_{\text{MC}}, rms (max)closed form's own residualMonte Carlo noise12σ2S2⋅∂2V/∂S2\tfrac12\sigma^2 S^2 \cdot \partial^2 V/\partial S^2 term, rms
τ=1\tau = 11.70.044 (0.083)0.00160.0515.19
τ=12\tau = \tfrac120.30.026 (0.056)0.00330.0336.27

Together these checks complete the circle. The discounted stock behaves as a martingale under QQ, not under the real-world drift, and the risk-neutral expectation generates the same pricing function that solves 2.7's PDE. Feynman–Kac's statement, that the expectation and the PDE are two descriptions of one object, is not only derived in the theory section but visible in the numbers themselves.

Into 2.9

2.8 leaves pricing in the form of an expectation. The simulator estimates

V=e−rT⋅EQ[payoff],V = e^{-rT} \cdot \E^Q[\text{payoff}],

the PDE accepts the same function as its solution, and the experiments show that the two routes arrive at the same number. The remaining question is whether that expectation can be evaluated directly rather than by drawing paths.

For this particular payoff, it can. Under the risk-neutral measure the terminal stock price is lognormal, so the price becomes an ordinary integral: the discounted average of max⁡(ST−K,0)\max(S_T - K, 0) against the lognormal density. The Monte Carlo estimator of this page and the closed form used as its benchmark are therefore evaluating the same expectation in two different ways, one by sampling and one by calculus.

The next page performs that calculation. The integral is rearranged into pieces that can be expressed using the standard normal distribution, and the familiar quantities N(d1)N(d_1) and N(d2)N(d_2) emerge naturally from the algebra. Nothing new is being priced and no new assumptions are introduced. The risk-neutral expectation established here is simply carried through to a closed-form answer.

The result will be the formula that this page treated as a benchmark: the Black–Scholes call price. Here it was accepted because 2.7 showed that it satisfies the PDE; in 2.9 it will be derived directly from the expectation itself. The Monte Carlo pricer, the expectation and the closed form will then be three views of the same value rather than three separate ideas.

Check

You price a European call by Monte Carlo: draw S_T many times, average the discounted payoff e^(−rT)·max(S_T − K, 0). The drift you simulate with is:

Why does a drift the stock does not have give the right price?

Under the r-drift, the discounted stock price e^(−rt)S_t is:

You have 10⁶ paths simulated at the stock’s real-world drift μ = 0.3. To price under Q you must:

Your pricer reports 18.35 ± 0.31 from 10⁴ paths; the closed form is 18.02. The pricer is: