How Girsanov Change-of-Measure Algorithms Turn Real-World Drifts into Risk-Neutral Pricing: Radon–Nikodym Weights, Market Price of Risk, Martingales and Failure Cases

Reader question: A stock can have an expected real-world return of 8%, 12% or something else, yet an option-pricing model often replaces that drift by the risk-free rate. Where did the original expected return go?

Girsanov’s theorem explains the change without pretending the physical drift never existed. The theorem changes the probability weighting of paths. Under the new measure, a drift-adjusted process becomes Brownian motion, while the diffusion scale remains the same. In finance, this is the machinery that connects a physical probability model to an equivalent martingale measure used for arbitrage-free pricing.

This article owns the Brownian change-of-measure problem: physical drift + diffusion exposure + market price of risk + density-process conditions → a new probability measure under which discounted tradable prices have the required martingale property.

It does not own historical return forecasting, option-volatility calibration, the estimation of risk premia, or every possible jump-process measure change. Those are separate jobs. The purpose here is to make the change of measure computable, testable and conceptually precise.

This is public mathematical education. It is not investment advice, and “risk-neutral” does not mean investors are actually indifferent to risk.

1. Two probability measures can describe the same possible paths

Let P denote a physical or real-world probability measure. It is intended to describe how likely outcomes are believed to be under a chosen statistical model.

Let Q denote a pricing measure.

If Q is equivalent to P, the two measures agree on which events have probability zero. A path that is possible under one is not made impossible under the other. What changes is the weight assigned to the path.

This distinction matters. Measure change is not path deletion. It is probability reweighting.

2. Begin with one Brownian stock under P

For the simplest example, suppose a non-dividend-paying stock follows:

dS_t/S_t = μdt + σdW_t^P.

Here:

• μ is the physical expected return;
• σ > 0 is volatility;
• W^P is Brownian motion under P.

Let the money-market account grow at rate r.

For arbitrage-free risk-neutral pricing in the standard complete Black–Scholes setting, the discounted stock:

S_te^−rt

should be a martingale under Q.

3. The market price of Brownian risk is the drift gap per unit volatility

Define:

λ = (μ−r)/σ.

This is the amount of physical excess drift per unit diffusion exposure in this one-factor example.

Now define a shifted process:

W_t^Q = W_t^P + λt.

Girsanov’s theorem tells us that, under a properly constructed equivalent measure Q, this shifted process is Brownian motion.

Since:

dW^P = dW^Q − λdt,

substitute into the stock equation:

dS/S = μdt + σ(dW^Q − λdt).

Therefore:

dS/S = (μ−σλ)dt + σdW^Q.

By construction:

μ−σλ = r.

So under Q:

dS/S = rdt + σdW^Q.

The physical drift has not been “proved wrong.” The probability measure has changed.

4. The Radon–Nikodym density tells us how paths are reweighted

For constant λ over a finite horizon T, the density process takes the exponential-martingale form:

Z_T = exp[−λW_T^P − ½λ²T].

Then Q is defined by:

dQ/dP = Z_T.

For a time-varying adapted process λ_t, the corresponding stochastic exponential is:

Z_T = exp[−∫₀^T λ_sdW_s^P − ½∫₀^T λ_s²ds].

The density must be a valid positive martingale with expectation one for this construction to define the intended equivalent probability measure.

5. Novikov’s condition is a sufficient gate, not the theorem itself

A widely used sufficient condition is Novikov’s condition:

E^P[exp(½∫₀^T λ_s²ds)] < ∞.

When it holds, the stochastic exponential is a true martingale under the standard setup.

But Novikov is sufficient, not necessary. Failing a numerical Novikov check does not automatically prove that no valid measure change exists. It means that this convenient sufficient route did not certify it.

This distinction is important in model validation: a diagnostic gate should not be exaggerated into an equivalence theorem.

6. Why volatility does not disappear under the Brownian measure change

Notice what changed in:

dS/S = μdt + σdW^P

versus:

dS/S = rdt + σdW^Q.

The drift changed. The local diffusion coefficient σ did not.

For a Brownian diffusion under an equivalent Girsanov change of measure, the theorem changes drift through the Brownian shift; it does not simply let us choose a new volatility process arbitrarily.

This is one reason volatility remains central to option pricing while the physical stock drift disappears from the Black–Scholes price.

7. Dividends change the target drift, not the logic

For a stock paying continuous dividend yield q, the risk-neutral ex-dividend price dynamics are:

dS/S = (r−q)dt + σdW^Q.

If the physical drift of the ex-dividend price is μ, the one-factor market price of risk becomes:

λ = [μ−(r−q)]/σ.

The same machinery works: identify the drift required by the chosen numeraire and tradable structure, then use a measure change that removes the excess physical drift.

8. Risk-neutral probabilities are pricing probabilities, not forecasts

This is one of the most important safety boundaries in mathematical finance.

Suppose Q assigns a 30% probability to an event and P assigns 15%. It does not follow that the event “really” has probability 30%.

Q is selected so discounted tradable prices satisfy martingale restrictions. Its probabilities combine physical likelihoods and risk pricing.

Using Q to answer a real-world forecasting question can therefore be a category error.

Likewise, using P directly to discount a derivative payoff without incorporating risk preferences or replication can violate arbitrage pricing.

9. The Fundamental Theorem of Asset Pricing supplies the larger structure

In modern arbitrage theory, under suitable technical conditions, absence of arbitrage in a market is linked to the existence of an equivalent martingale measure.

In finite or suitably regular continuous-time settings, the intuition is:

no arbitrage → there exists a probability measure under which discounted tradable prices behave as martingales.

Market completeness then relates to uniqueness of that pricing measure.

Girsanov’s theorem is not identical to the Fundamental Theorem of Asset Pricing. Girsanov gives a way to transform Brownian-driven dynamics under a change of measure. The asset-pricing theorem tells us why an equivalent martingale measure matters economically.

10. Multidimensional markets turn λ into a vector

Suppose a vector of risky assets has diffusion matrix Σ and physical excess-drift vector b. A Brownian market-price-of-risk vector λ must satisfy a relation of the form:

b = Σλ.

If there is a unique λ compatible with all traded risks, the diffusion market can be complete under suitable regularity conditions.

If multiple λ vectors satisfy the pricing restrictions because some Brownian risks are not spanned by traded assets, the market is incomplete and no-arbitrage alone may not determine a unique price for every non-replicable claim.

If no λ solves the required drift equations in a supposed frictionless diffusion model, the specification is inconsistent with the desired equivalent-martingale-measure condition.

11. Stochastic volatility exposes incompleteness

In a simple Heston-style model, one Brownian motion drives the asset and another partly independent Brownian source drives variance.

If variance risk cannot be perfectly traded away using available securities, the market is incomplete. The physical-to-pricing transformation then needs a choice about the price of volatility risk.

This is why moving from P to Q in stochastic-volatility models is not generally as simple as replacing μ by r. Some risk-premium component is not identified by no-arbitrage alone.

Calibration to option prices can identify a risk-neutral parameter set, but it does not automatically reveal the corresponding physical risk premium uniquely.

12. Change of numeraire is measure change with an economic purpose

The money-market account is not the only possible numeraire. A positive tradable asset can sometimes be used as the unit in which other prices are expressed.

If N_t is a valid positive numeraire, one can construct a corresponding measure Q^N under which prices divided by N are martingales.

Starting from a money-market numeraire B, a standard density relation is proportional to:

(N_t/B_t)/(N₀/B₀).

This technique can simplify interest-rate, FX, quanto and credit calculations by choosing the numeraire that makes the target payoff’s expectation easiest to compute.

13. Forward measures are a practical example

Take a zero-coupon bond maturing at T as numeraire. Under the associated T-forward measure, appropriately normalised asset prices become martingales relative to that bond.

This often removes discounting from an expectation of a payoff paid at T.

The algorithmic idea is:

choose numeraire → derive density process → transform Brownian drift → evaluate the payoff under the measure that simplifies its dynamics.

Changing measure is therefore not merely a proof trick. It can be a computational design choice.

14. Inputs and outputs of a change-of-measure engine

Inputs can include:

• physical drift vector;
• diffusion matrix;
• risk-free or numeraire dynamics;
• dividend/funding/carry conventions;
• candidate market-price-of-risk process;
• time horizon;
• integrability assumptions;
• traded-asset set defining pricing constraints.

Outputs can include:

• candidate λ process;
• Radon–Nikodym density process;
• Q-drift of each state variable;
• discounted-price martingale residuals;
• existence/uniqueness diagnostics;
• numeraire-adjusted dynamics;
• Monte Carlo likelihood weights where measure reweighting is used computationally.

15. A minimal one-factor algorithm

1. Write the P-dynamics of the tradable.
2. Write the drift required under the pricing numeraire.
3. Compute the drift gap.
4. Divide by diffusion exposure to obtain λ in the one-factor case.
5. Construct the stochastic-exponential density.
6. Verify a sufficient martingale/integrability condition or otherwise establish validity.
7. Define Q through the density.
8. Substitute dW^P = dW^Q − λdt.
9. Verify discounted tradables have zero drift under Q.
10. Price replicable claims as discounted Q-expectations or by equivalent hedging/PDE methods.

16. Evidence polarity

Evidence for confidence includes a positive density process with expectation one, transformed drifts that exactly satisfy the numeraire-martingale restrictions, unchanged diffusion covariance where required, agreement between Q-expectation prices and independent replication/PDE formulas, stable Monte Carlo likelihood weights, and consistent results under equivalent numeraire transformations.

Evidence against confidence includes density weights whose sample mean drifts materially from one, explosive or highly concentrated likelihood weights, transformed discounted assets with residual drift, an unsolved multidimensional drift equation, multiple pricing measures presented as one unique truth, or use of risk-neutral probabilities as physical default/return forecasts without an explicit bridge back to P.

17. Counterexample: replacing μ by r without changing the probability law

A student can write a new SDE with drift r and call it “risk neutral” without defining any relationship between the old and new measures.

The resulting formula may resemble Black–Scholes, but the probabilistic step is missing.

Falsifier: identify the density process or equivalent replication argument that supports the change. If no measure or hedge relation exists, the drift replacement is an assertion rather than a derivation.

18. Counterexample: risk-neutral default probability mistaken for real default frequency

Credit derivatives can imply risk-neutral hazard rates. Those rates include compensation for priced credit risk and liquidity/model effects.

Using them directly as forecasts of realised default frequency confuses Q with P.

Falsifier: compare model-implied Q probabilities with historical/default-forecast models and explicitly identify the risk-premium bridge required to move between them.

19. Counterexample: incomplete markets create multiple Q measures

If an untraded volatility factor introduces an extra Brownian source, many choices of its market price of risk can satisfy the stock’s no-arbitrage condition.

A single derivative price may then require additional modelling, calibration or hedging assumptions.

Falsifier: count independent sources of risk and the rank of traded diffusion exposures. If some risk directions are unspanned, uniqueness must be justified rather than assumed.

20. Counterexample: a stochastic exponential can be a local martingale but not a valid density

Writing down:

Z = exp(stochastic integral − half quadratic variation)

does not by itself guarantee E[Z_T] = 1.

In pathological or aggressive specifications, the exponential local martingale can fail to be a true martingale.

Falsifier: establish the required martingale property analytically or through a justified theorem; numerical evidence of severe weight loss is a warning, not a proof.

21. Counterexample: likelihood-ratio simulation can have terrible variance

Measure changes are also used for importance sampling. One deliberately changes the path distribution to make rare events more common, then multiplies by likelihood ratios to restore the original expectation.

If the chosen measure is poor, a few paths receive enormous weights and dominate the estimate.

Falsifier: inspect effective sample size, weight concentration and repeated-run variance. An unbiased identity can still be computationally useless.

22. Counterexample: volatility cannot be changed freely by Brownian Girsanov

A proposed “measure change” that transforms σ from 20% to 35% while claiming only to apply the standard Brownian drift theorem is not the same Girsanov operation described here.

Falsifier: write both SDEs and compare quadratic variation. Equivalent Brownian drift changes preserve the local diffusion structure under the standard theorem.

23. Alternatives and neighbouring tools

Replication can derive derivative values without beginning from probability measure language; in complete markets it leads to the same price.

Feynman–Kac/PDE methods convert the risk-neutral dynamics into a differential equation.

Esscher transforms are often used for certain Lévy models and change exponential tilting in a way adapted to jump distributions.

Minimal martingale, entropy and utility-based measures are examples of selection principles in incomplete markets.

Change of numeraire is a specialised measure-change strategy chosen to simplify a particular pricing problem.

24. Verification ladder

1. Drift equation: verify that the proposed λ exactly removes the required excess drift.
2. Density positivity: verify Z is positive where equivalence is claimed.
3. Normalisation: establish E[Z]=1 under the chosen assumptions.
4. Martingale test: simulate or analytically verify discounted tradables have zero conditional drift under Q.
5. Quadratic-variation test: confirm the diffusion structure is transformed correctly.
6. Replication benchmark: compare with Black–Scholes/PDE or another independent pricing route.
7. Rank test: inspect the multidimensional diffusion matrix for completeness/uniqueness issues.
8. Numeraire round-trip: price the same claim under two valid numeraires and reconcile results.
9. Weight diagnostic: if reweighting paths, inspect variance and effective sample size.
10. P-versus-Q boundary: label every probability output by the measure under which it is defined.

25. Connections to the surrounding Bukit Timah Tutor knowledge estate

The simplest risk-neutral endpoint is developed in Black–Scholes option-pricing algorithms. This page explains the probability-measure machinery beneath the familiar replacement of physical drift by the pricing drift.

Forward-curve models use the same no-arbitrage logic at much larger scale. Heath–Jarrow–Morton algorithms show how risk-neutral no-arbitrage constrains the drift of an entire forward-rate curve once volatility is specified.

Short-rate calibration connects through Hull–White calibration algorithms.

Foreign-exchange no-arbitrage connects to FX-forward pricing algorithms, where domestic and foreign numeraires and carry relationships determine forward prices.

Monte Carlo implementation connects to Monte Carlo pricing algorithms and to importance-sampling algorithms, where likelihood ratios turn change-of-measure theory into a variance-reduction tool.

The full lane is indexed at Finance & Banking Algorithms | Applied Mathematics in Real Financial Systems.

26. What would falsify confidence?

Confidence should be withdrawn if the density process cannot be justified as a true martingale; discounted tradables retain systematic drift under the supposed pricing measure; a multidimensional drift system has no solution; an incomplete model is presented as having a unique Q without additional assumptions; prices disagree with independent replication; or outputs labelled as “probabilities” silently mix P and Q.

27. Verification and update triggers

Preserve the P-dynamics, Q-dynamics, diffusion matrix, λ definition, density process, integrability assumptions, numeraire, calibration/risk-premium choices and martingale tests for every model version.

Revalidate when state variables, tradable assets, funding/dividend conventions, numeraire, correlation structure, volatility factors or measure-selection assumptions change. Trigger review when likelihood weights become unstable, new unspanned risk factors are introduced, martingale residuals appear, or a model previously used only for pricing is repurposed for physical forecasting.

28. Primary and high-quality references

• Walter Schachermayer, Fundamental Theorem of Asset Pricing, Wiley reference overview.
• Patrick S. Hagan and Andrew Lesniewski, Girsanov, Numeraires, and All That, Cambridge University Press, 2022.
• University of Vienna, Mathematical Finance Lecture Notes, section on Girsanov’s theorem and mathematical-finance applications.
• MIT OpenCourseWare, Analytics of Finance — arbitrage-free pricing lecture material.
• Andrew Lesniewski, Baruch College, Interest Rate Models: Girsanov, Numeraires, and All That.

Educational boundary: Girsanov changes probability weights so that selected stochastic dynamics acquire the required drift under a new measure. It does not tell us which physical forecasts are true, and in incomplete markets it does not by itself select a unique price for every unhedgeable claim.