Risk-Neutral Pricing: Why Derivatives Don't Depend on Expected Returns

Here is one of the most surprising facts in mathematical finance: the price of a European call option on a non-dividend-paying stock does not depend on the expected return of the stock. It depends on the risk-free rate, the volatility, the strike, the maturity, and the current price. The expected return is irrelevant. You can change your mind about whether stocks earn 8% or 12% per year, and the Black-Scholes price of a call does not change.

This is not an accident. It is the deep content of the fundamental theorem of asset pricing. The same fact underlies Monte Carlo pricing, exotic option valuation, and most of quantitative finance. The mechanism is a change of measure called Girsanov’s theorem. This article derives both, explains why option prices depend on volatility but not drift, and shows the practical implications.

The Two Measures

There are two measures over the same probability space that matter:

• The physical measure $\mathbb{P}$, under which expected returns equal what investors expect. Under $\mathbb{P}$, the stock earns $\mu$, which is some unknown quantity estimated from data.
• The risk-neutral measure $\mathbb{Q}$, under which every discounted asset price is a martingale. Under $\mathbb{Q}$, the stock earns $r$, the risk-free rate.

Both measures are equivalent in a precise technical sense (they agree on null sets), but they weight future outcomes differently. Specifically, $\mathbb{Q}$ re-weights outcomes by their marginal utility to a representative risk-averse investor. Outcomes that are bad in utility terms (like a market crash) get lower probability under $\mathbb{Q}$ than under $\mathbb{P}$, because the investor would be willing to pay insurance against them.

The key pricing identity is

for any tradable asset or derivative $V$. This says: discount the expected payoff under $\mathbb{Q}$. The expected return $\mu$ does not appear. The risk premia are baked into the measure $\mathbb{Q}$.

Why Martingales Matter

A process $M_t$ is a martingale under $\mathbb{Q}$ if $\mathbb{E}^{\mathbb{Q}}[M_T | \mathcal{F}_t] = M_t$. The intuitive reading is: given what you know now, your best forecast of the future value of $M$ is its current value. Martingales are the right notion of “fair game” in a stochastic setting.

If $S_t$ is a stock price and $r$ is the risk-free rate, then the discounted stock price $\tilde{S}_t = e^{-rt} S_t$ is a martingale under $\mathbb{Q}$. This is exactly the no-arbitrage condition: the discounted stock cannot have predictable growth, because if it did, you could trade on that predictability and make risk-free profit. Under $\mathbb{Q}$, the expected growth rate of $S_t$ is exactly $r$. The expected growth under $\mathbb{P}$ is higher (because stocks are risky), but the risk-adjusted expected growth is $r$.

For a derivative $V_t$ with payoff $V_T$ at time $T$, no-arbitrage also forces $\tilde{V}_t = e^{-rt} V_t$ to be a martingale. So

Set $t = 0$ and you get the pricing formula. This is the entire content of risk-neutral pricing: derivatives are discounted expectations under $\mathbb{Q}$.

The Fundamental Theorem

The fundamental theorem of asset pricing (FTAP) comes in two parts.

FTAP, first version. A market has no arbitrage if and only if there exists an equivalent martingale measure (a measure $\mathbb{Q}$ equivalent to $\mathbb{P}$ under which all discounted asset prices are martingales).

FTAP, second version. If the market is complete (every contingent claim can be replicated), the equivalent martingale measure is unique.

These theorems are the formal backbone of mathematical finance. They were proven in stages by Harrison, Kreps, Pliska (1979), Ross (1976), and others, building on the no-arbitrage arguments of Black, Scholes, and Merton.

For our purposes, the punchline is: no arbitrage $\Leftrightarrow$ there exists a risk-neutral measure. Pricing derivatives reduces to computing expectations under that measure.

Girsanov’s Theorem: The Heart of the Change

How do you go from $\mathbb{P}$ to $\mathbb{Q}$? The answer is Girsanov’s theorem. We state a simplified version for one Brownian motion.

Girsanov’s Theorem. Let $W_t$ be a standard Brownian motion under $\mathbb{P}$. Let $\theta_t$ be an adapted process satisfying the Novikov condition $\mathbb{E}^{\mathbb{P}}[\exp(\frac{1}{2} \int_0^T \theta_t^2 \, dt)] < \infty$. Define

and let $\mathbb{Q}$ be the measure with Radon-Nikodym derivative

Then $W_t^{\mathbb{Q}}$ is a standard Brownian motion under $\mathbb{Q}$.

The theorem says: if you shift a Brownian motion by an adapted drift, you get another Brownian motion under a tilted measure. The tilt — the density process — is the exponential martingale that defines $\mathbb{Q}$.

For our stock-price SDE under $\mathbb{P}$,

the idea is to choose $\theta$ such that the drift becomes $r$ under $\mathbb{Q}$. Solving $\mu - \sigma \theta = r$ gives $\theta = (\mu - r)/\sigma$, the market price of risk (Sharpe ratio divided by volatility). Define

Then under $\mathbb{Q}$,

That is the SDE under $\mathbb{Q}$: GBM with drift $r$. The volatility $\sigma$ is unchanged — only the drift changes. This is the change of measure in action.

The Market Price of Risk

The quantity $\theta = (\mu - r)/\sigma$ is called the market price of risk for the stock. It measures how much extra return per unit of volatility investors demand for holding the stock. Equivalently, it is the Sharpe ratio divided by the volatility — the trade-off between risk and reward that the market is willing to make.

In a multi-asset world, each asset has its own market price of risk. They are not independent — they must satisfy certain consistency conditions so that arbitrage does not appear. For $n$ assets driven by $d$ Brownian motions, you have $n$ market prices of risk and $d$ random sources, giving an $n \times d$ matrix that determines the risk premia. As long as this matrix is consistent with a valid equivalent measure, the market is arbitrage-free.

The existence of a market price of risk for every asset is equivalent to the existence of an equivalent martingale measure. It is the same statement in different language.

A Complete Derivation for GBM

Let us redo the change of measure carefully. The setup:

• $W_t$ is a Brownian motion under $\mathbb{P}$.
• $S_t$ satisfies $dS_t = \mu S_t \, dt + \sigma S_t \, dW_t$ under $\mathbb{P}$.
• We want a measure $\mathbb{Q}$ under which $dS_t = r S_t \, dt + \sigma S_t \, dW_t^{\mathbb{Q}}$.

Define the process

(constant, since $\mu$, $r$, $\sigma$ are constant here). Define

and the density process

By Girsanov, $W_t^{\mathbb{Q}}$ is a Brownian motion under $\mathbb{Q}$. Substituting $dW_t = dW_t^{\mathbb{Q}} - \theta \, dt$ into the stock SDE:

because $\sigma \theta = \mu - r$. Done. The change of measure is exact, and under $\mathbb{Q}$ the stock earns the risk-free rate.

This is more than a mathematical trick. It says: there exists a probability measure under which the stock grows at the risk-free rate. The measure is constructed explicitly. Discounting the expected payoff under this measure gives the right price.

The Black-Scholes Formula via Risk-Neutral Pricing

Given $dS_t = r S_t \, dt + \sigma S_t \, dW_t^{\mathbb{Q}}$ with $S_0$ given, the explicit solution is

For a European call with payoff $(S_T - K)^+$, the price is

Computing the expectation with the lognormal distribution of $S_T$ gives the Black-Scholes formula:

The expected return $\mu$ does not appear. The risk-free rate $r$ appears. This is the cleanest demonstration of risk-neutral pricing: a formula with no $\mu$ in it, derived from an SDE that contains $\mu$.

Hedging and Completeness

The other side of the FTAP is completeness. A market is complete if every contingent claim can be replicated by a trading strategy. In the Black-Scholes setup, the market is complete. There is a unique equivalent martingale measure, and the hedging strategy is the famous delta hedge: hold $\partial C / \partial S$ shares of stock per option sold.

If the market were not complete — for example, if there were multiple sources of randomness and not enough traded assets to span them — then there would be multiple equivalent martingale measures, and the price of a derivative would not be uniquely determined by no-arbitrage alone. There would be a range of no-arbitrage prices. The width of the range measures the incompleteness.

In practice, most markets are approximately complete. Stocks span the volatility of stocks. Options span the volatility of options. There are exotic claims that cannot be exactly replicated (deep out-of-the-money options on illiquid underlyings), but those are corner cases.

Monte Carlo Pricing

The pricing formula $V_0 = e^{-rT} \mathbb{E}^{\mathbb{Q}}[V_T]$ is a recipe. To price a derivative by Monte Carlo, you:

1. Simulate paths of the underlying under $\mathbb{Q}$ (drift $r$, volatility $\sigma$, etc.).
2. Compute the payoff $V_T$ on each path.
3. Average the discounted payoffs.

This is the entire algorithm. You do not simulate under $\mathbb{P}$. You do not use $\mu$. You use the risk-neutral SDE.

For an exotic payoff like an Asian option, there is no closed form, and Monte Carlo is the standard approach. For a barrier option, you might use a closed-form approximation or simulate. For a basket option, simulation is essentially the only option beyond moment-matching approximations.

import numpy as np

def price_european_call(S0, K, r, sigma, T, n_paths=200000, seed=42):
    rng = np.random.default_rng(seed)
    Z = rng.standard_normal(n_paths)
    ST = S0 * np.exp((r - 0.5 * sigma**2) * T + sigma * np.sqrt(T) * Z)
    payoff = np.maximum(ST - K, 0.0)
    return np.exp(-r * T) * payoff.mean()

print(price_european_call(100, 100, 0.05, 0.20, 1.0))

The output should be close to the Black-Scholes value of 10.45 for these inputs. Run with more paths for tighter error bars. Note that $\mu$ does not appear anywhere. Change it and the price does not move.

Why Option Prices Are Sensitive to Volatility, Not Drift

The intuition: an option’s payoff depends on the distribution of the terminal price, not its expected value. The expected value of the terminal price is $S_0 e^{rT}$ regardless of $\mu$ — under $\mathbb{Q}$, the drift is $r$, not $\mu$. So changing $\mu$ does not change the expected terminal price, just the measure under which you compute expectations.

But changing $\sigma$ changes the entire distribution. Higher volatility means fatter tails, more extreme outcomes, more chance of finishing deep in-the-money for a call. So call prices go up with vol. This is the leverage effect, and it is why vega (the sensitivity to vol) is one of the most important Greeks.

The same argument explains why out-of-the-money puts get expensive in volatile markets: the fatter left tail of the lognormal under high vol makes deep losses more likely. Risk-neutral pricing makes this precise.

The Limits of Risk-Neutral Pricing

Risk-neutral pricing works when:

• The market is arbitrage-free.
• The payoff is attainable (or approximately so).
• The risk-neutral measure is unique or known.

It does not work when:

• The market is illiquid and bid-ask spreads dominate (price is no longer a clean expectation).
• The payoff is highly path-dependent and the model is wrong (you get an arbitrage-free price that nobody will trade at).
• The model has fat tails or jumps and the option is far out-of-the-money (the model price is way off from what a real bank would quote).

For practical work, risk-neutral pricing is a baseline. You compute the no-arbitrage price, then compare to market quotes. The gap is your model’s error. If the gap is small, you trust the model. If the gap is large, you need a better model.

What to Take Away

The fundamental theorem of asset pricing says no arbitrage $\Leftrightarrow$ existence of equivalent martingale measure $\mathbb{Q}$. Girsanov’s theorem gives you the construction: shift the Brownian motion by the market price of risk, and you get a $\mathbb{Q}$-Brownian motion. Under $\mathbb{Q}$, the stock grows at $r$, and the price of a derivative is the discounted $\mathbb{Q}$-expectation of the payoff.

The expected return $\mu$ disappears. Volatility remains. This is the cleanest expression of the fact that derivative pricing is about risk, not return.

References: Björk (2009), Arbitrage Theory in Continuous Time, Chapters 10-11; Shreve (2008), Stochastic Calculus for Finance II, Chapter 5; Hull (2018), Ch. 15.7 (risk-neutral pricing in the Black-Scholes-Merton chapter). Glasserman (2004) discusses Monte Carlo under $\mathbb{Q}$ throughout. The conceptual punchline is simple. The technical detail takes a chapter or two. Take your time with it.

The Radon-Nikodym Derivative, Carefully

The density process $Z_t = d\mathbb{Q}/d\mathbb{P}|_{\mathcal{F}_t}$ is an exponential martingale. Its dynamics are

so $Z_t = \exp(-\theta W_t - \frac{1}{2} \theta^2 t)$. The Novikov condition $\mathbb{E}^{\mathbb{P}}[\exp(\frac{1}{2} \int_0^T \theta^2 \, dt)] < \infty$ ensures that $Z_T$ has expectation 1, which is the requirement for a probability density.

For constant $\theta$, the Novikov condition becomes $\exp(\frac{1}{2} \theta^2 T) < \infty$, which is always true. So for constant market price of risk, the change of measure is well-defined for all $T$. For stochastic $\theta_t$, you need additional conditions to ensure integrability.

The exponential martingale has a nice property: it is the unique martingale $Z_t$ with $\mathbb{E}^{\mathbb{P}}[Z_T \cdot X_T] = \mathbb{E}^{\mathbb{Q}}[X_T]$ for any claim $X_T$. So if you can find the right density, you can compute expectations under $\mathbb{Q}$ by computing expectations under $\mathbb{P}$ weighted by $Z_T$. This is the basis of importance sampling, a Monte Carlo variance reduction technique.

The Cameron-Martin Space

Girsanov’s theorem has a geometric interpretation. The space of equivalent Brownian motions under equivalent measures is the Cameron-Martin space — the closure of absolutely continuous functions under the $L^2$ norm. The drift $\theta_t$ that appears in Girsanov’s theorem must lie in this space for the change of measure to be valid.

For a process $\theta_t$ outside the Cameron-Martin space (e.g., a process with too much volatility of its own), no equivalent measure exists that makes the shifted Brownian motion a true Brownian motion. This is a regularity condition. In most finance applications, $\theta_t$ is bounded or at most linearly growing, so we are safely inside the Cameron-Martin space.

The Multi-Dimensional Case

When there are multiple Brownian motions and multiple assets, the change of measure generalizes. Let $\mathbf{W}_t = (W_t^1, \ldots, W_t^d)$ be a vector of independent Brownian motions under $\mathbb{P}$. Let the $i$-th asset satisfy

Define a $d$-dimensional process $\boldsymbol{\theta}_t = (\theta_t^1, \ldots, \theta_t^d)$ — the vector of market prices of risk — by requiring

for each $i$. This is a linear system $n \times d$ where $n$ is the number of assets. If $d \geq n$ and the volatility matrix has full row rank, $\boldsymbol{\theta}$ is determined uniquely. If $d < n$, the system is underdetermined and we have multiple risk-neutral measures — the market is incomplete.

Under $\mathbb{Q}$,

where $\mathbf{W}_t^{\mathbb{Q}} = \mathbf{W}_t + \int_0^t \boldsymbol{\theta}_u \, du$. The density process is

This is the multi-asset Girsanov theorem. For Black-Scholes with one asset and one Brownian motion, it reduces to the scalar case we derived above.

The Bayes Interpretation

There is a beautiful Bayesian reading of the change of measure. Under $\mathbb{P}$, you have a prior over market scenarios weighted by their physical likelihood. Under $\mathbb{Q}$, you have a posterior weighted by their economic value. The density $Z_T$ is the likelihood ratio that takes you from prior to posterior.

This is not a metaphor. It is a precise statement. The Radon-Nikodym derivative $Z_T$ is the Bayes factor. Pricing under $\mathbb{Q}$ is Bayesian inference with a particular choice of prior. Risk premia are the costs of updating from physical to risk-neutral beliefs.

The implication: derivative prices encode information about risk preferences. Out-of-the-money put prices being expensive (relative to lognormal) means the market is pricing in a high cost of equity drawdowns. This is the foundation of risk-neutral inference: from option prices, back out the implied risk-neutral distribution, and read off market risk premia. That is the entire field of model-free implied volatility and risk-neutral moments.

When Risk-Neutral Pricing Fails in Practice

Three scenarios where the formula breaks:

1. Default risk. If the counterparty can default, the payoff $(S_T - K)^+$ is not guaranteed. You have to discount by the survival probability and adjust for recovery. The result is a credit-adjusted price, no longer a clean expectation under $\mathbb{Q}$.

2. Liquidity risk. If you cannot trade at the mid price (you have to cross the spread), the no-arbitrage argument breaks. Prices depend on liquidity premia, not just discounted expectations.

3. Funding constraints. If your hedging program is subject to funding costs (collateral, repo, margin), the price of hedging is not just the option payoff. There is a funding cost that varies over time and creates a residual basis.

For vanilla products on liquid underlyings, these effects are small. For exotics, structured products, and distressed names, they dominate. Many real-world derivatives trade at prices that are far from the textbook risk-neutral expectation.

A Subtle Point: The Numeraire

The risk-neutral measure is not unique. It depends on which asset you choose as the numeraire — the unit of account in which you discount. The measure associated with the risk-free bank account $B_t = e^{rt}$ as numeraire is $\mathbb{Q}^B$, often just called $\mathbb{Q}$. But you can equally well define a measure $\mathbb{Q}^S$ associated with the stock $S_t$ itself as numeraire. Under $\mathbb{Q}^S$, $1/S_t$ discounted is a martingale, which means $S_t$ itself has zero drift under $\mathbb{Q}^S$.

Different numeraires give different risk-neutral measures, but they all give the same price for any tradable claim. This is numeraire invariance: the price of a derivative is the same whether you compute it under $\mathbb{Q}^B$ or $\mathbb{Q}^S$.

The practical use of numeraire invariance: if you can change to a numeraire under which an expectation is easier to compute, do so. For example, pricing under the stock numeraire turns a call option’s expectation into a probability question about whether the stock finishes above the strike. This gives the digital option price in one line.

An Aside on Brownian Motion Under $\mathbb{Q}$

A subtle but important point: a process being a “Brownian motion” is a measure-dependent statement. The same process $W_t$ is a Brownian motion under $\mathbb{P}$ but $W_t + \theta t$ is a Brownian motion under $\mathbb{Q}$. Both are continuous martingales with quadratic variation $t$. Both have Gaussian increments. But only one of them has mean zero increments under each measure.

This is why the notion of “Brownian motion” in finance is always measure-qualified. We say “Brownian motion under $\mathbb{Q}$” to be precise. The same $W_t$ under different measures is a different object, with different dynamics.