Martingale Measures

Modern derivatives pricing rests on a single, powerful idea: in a frictionless market populated by rational agents, no asset can have two prices simultaneously. If it did, traders would instantaneously exploit the difference and competitive pressure would eliminate the discrepancy. From this deceptively simple observation, the absence of arbitrage, we can derive a complete, internally consistent framework for pricing any contingent claim, from a vanilla European call to an exotic path-dependent option.

The theory is one of the most beautiful in quantitative finance. It is also, in important ways, wrong or more precisely, contingent on a set of idealised assumptions that the real world routinely violates. Understanding both the power of the theory and the precise mechanisms of its failure is essential for anyone working in markets.

This blog builds the theory from the ground up, starting with the no-arbitrage principle, martingales and Girsanov’s theorem, and then carefully going through each assumption to understand what happens when markets are not frictionless.

The no-arbitrage principle

An arbitrage is a trading strategy that requires zero initial investment, carries no risk of loss, and generates a strictly positive profit with positive probability. More precisely, a self-financing portfolio strategy φ is an arbitrage if:

That is: start with nothing, end with something non-negative almost surely, and have a genuine chance of ending with something strictly positive. The no-arbitrage condition is simply the requirement that no such strategy exists. In any well-functioning market with rational, profit-seeking agents, arbitrage opportunities should be competed away almost instantaneously.

The law of one price

The law of one price is a direct corollary of the no-arbitrage condition. It states that any two assets or portfolios that produce identical cash flows in every possible state of the world must trade at the same price today. If they did not, one could buy the cheaper and sell the costlier, earning a riskless profit.

If two self-financing portfolios A and B satisfy $V_T(A) = V_T(B)$ almost surely under all states of the world, then $V_t(A) = V_t(B)$ for all $t ∈ [0, T]$. Violation implies arbitrage.

The law of one price is the foundation on which all derivative pricing is built. A derivative contract is priced by identifying a portfolio of more primitive securities (the underlying, cash bonds) that replicates the derivative’s payoff in every state. The derivative must then be worth exactly what that replicating portfolio costs because if it were cheaper or more expensive, an arbitrage would exist.

Martingales

In probability theory, a stochastic process {$M_t$} adapted to a filtration {$F_t$} is a martingale under a probability measure Q if it satisfies the tower property of conditional expectations:

For all s ≤ t,

Intuitively, a martingale is a “fair game” process: given all current information, the best forecast of any future value is today’s value. The process has no predictable drift. This is the key property we need for arbitrage-free pricing, because in an efficient market, asset prices discounted at the risk-free rate should have no systematic drift, if they did, there would be a riskless expected excess return, which would be a form of arbitrage.

The fundamental theorem of asset pricing

Theorem: A market is arbitrage-free if and only if there exists at least one equivalent martingale measure (EMM) Q ~ P under which all discounted asset prices are martingales. The market is complete (every contingent claim can be replicated) if and only if the EMM is unique.

The measure Q is called the risk-neutral measure or equivalent martingale measure. It is equivalent to the real-world measure P in the technical sense that they agree on which events have probability zero. But it is a different measure, it reassigns probabilities so that all assets earn exactly the risk-free rate in expectation.

Discounted asset prices as martingales

Under the EMM Q, the price of every traded asset, discounted by the money-market account $B_t = e^{rt}$, is a martingale:

for all s ≤ t,

This is not a statement about the real world — it is a mathematical consequence of the no-arbitrage condition. In the real world under P, risky assets earn a risk premium above the risk-free rate. Under Q, that premium is stripped away, leaving only the risk-free drift.

Risk-neutral valuation

Given an arbitrage-free, complete market with a unique EMM Q, the price at time t of any derivative with payoff H at time T is uniquely determined by:

This is the central formula of derivatives pricing. It says: take the expected value of the discounted payoff, where the expectation is computed under the risk-neutral measure Q, not the real-world measure P. The measure does all the work, the drift of the underlying under Q is always the risk-free rate, regardless of what investors actually believe about expected returns.

Risk-neutral vs. real-world: a crucial distinction

This is one of the most commonly misunderstood aspects of the theory. Under Q, the stock price process follows:

Under P, the stock drifts at μ — the true expected return, which exceeds the risk-free rate by the equity risk premium. Under Q, the drift is replaced by r. The volatility σ is the same under both measures (the change of measure only affects the drift, not the diffusion coefficient, as we will see in the Girsanov theorem).

The risk-neutral measure does not say investors are indifferent to risk. It is a mathematical tool that absorbs risk preferences into the probability weights, allowing us to compute prices as simple expected values without ever modelling individual utility functions.

Risk-neutral valuation frees us from needing to know μ — the expected return on the risky asset — which is both unobservable and contested. All we need is r and σ. This is why derivatives can be priced with far greater precision than the underlying assets themselves.

Numeraire invariance

A powerful generalisation of risk-neutral pricing is that we can price relative to any positive, non-dividend-paying numeraire, not just the money-market account. This is the numeraire invariance principle. If we choose a numeraire $N_t$ and its associated measure $Q^N$, then any asset price $V_t$ satisfies:

This has enormous practical utility. When pricing interest rate options using forward measures, choosing the zero-coupon bond as numeraire eliminates the stochastic discounting factor, greatly simplifying the computation. The change between numeraires is governed by as we see next the Girsanov theorem.

Girsanov’s theorem and change of measure

We have established that pricing requires working under the risk-neutral measure Q rather than the real-world measure P. But Q is not given to us directly, we must construct it from P. The mathematical tool for changing probability measures in the context of stochastic processes is Girsanov’s theorem, one of the most important results in stochastic calculus.

The Radon-Nikodym derivative

Any change of equivalent probability measures can be represented by a Radon-Nikodym derivative, a positive random variable $Z_T$ that tells us how to re-weight probabilities from P to Q. For any event A:

The process {$Z_t = E^P[Z_T | F_t]$} is itself a martingale under P (by the tower property), and is strictly positive (since Q ~ P).

Girsanov’s theorem

Girsanov’s theorem specifies precisely how a Brownian motion transforms under a change of measure. Let $W_t^P$ be a Brownian motion under P, and let $θ_t$ be an adapted process (the market price of risk). Define the Radon-Nikodym derivative:

Under the regularity condition that $Z_t$ is a true martingale, Girsanov’s theorem states:

Define $W_t^Q = W_t^P + ∫₀ᵗ θ_s ds.$ Then $W_t^Q$ is a standard Brownian motion under the measure Q defined by $dQ/dP = Z_T.$ The change of measure shifts the drift of the Brownian motion by $θ_t$, while leaving the volatility structure unchanged.

Girsanov’s theorem underpins every change-of-numeraire calculation in interest rate modelling. Switching from the risk-neutral measure to the T-forward measure (using the T-maturity zero-coupon bond as numeraire) is a Girsanov transformation that removes the stochastic discounting, making the pricing of caplets and swaptions dramatically simpler.

The market price of risk and the drift shift

Under the real-world measure, the stock SDE is $dS = μ S dt + σ S dW^P.$ Writing $W^P = W^Q − ∫θ dt$ and substituting:

For S to be a martingale under Q (discounted by $e^{rt}$), we need the drift under Q to equal r. This requires:

The quantity $θ = (μ − r) / σ$ is the Sharpe ratio of the asset, the excess return per unit of volatility. It is also called the market price of risk. Girsanov’s theorem tells us that switching from the real-world to the risk-neutral measure is precisely equivalent to subtracting the market price of risk from the Brownian motion’s drift.

Replication strategies and perfect hedging

The risk-neutral pricing formula gives us the price of a derivative as an expectation. But this expectation is not just a number pulled from a model, it corresponds to a concrete trading strategy that replicates the derivative’s payoff exactly. This replication strategy is the heart of the Black-Scholes framework and the bridge between the abstract mathematics and the trading desk.

A replication strategy is a self-financing portfolio $φ = (Δ_t, B_t)$, holdings in the underlying and the risk-free bond, such that the portfolio’s value at any time equals the derivative’s theoretical price, and its terminal value exactly equals the derivative’s payoff:

Portfolio value:

Self-Financing:

Payoff match:

Delta hedging and the Black-Scholes PDE

In the Black-Scholes model, the replication strategy for a European option with value $V(S,t)$ requires holding exactly $Δ_t = ∂V/∂S$ shares of the underlying at each moment in time. This is the famous delta of the option. By Itô’s lemma:

The self-financing replication argument eliminates all randomness (the dW terms cancel when we hold Δ = ∂V/∂S units of the stock), leaving a deterministic equation. No-arbitrage then requires this deterministic portfolio to earn the risk-free rate, yielding the Black-Scholes PDE:

Crucially, μ does not appear, it was eliminated by the hedging argument. The PDE is solved backward from the terminal condition $V(S,T) = H(S_T)$, and the solution is precisely the risk-neutral expectation $E^Q[e^(−rT)H(S_T)]$. The PDE approach and the measure-theoretic approach are two sides of the same coin.

Dynamic delta hedging in continuous time

Perfect replication requires continuous rebalancing of the delta hedge. The procedure is:

• At each instant $t$, compute $Δ_t = ∂V/∂S$ => the sensitivity of the option value to the underlying price.
• Hold exactly $Δ_t$ units of the underlying in the portfolio. Finance this position by borrowing or lending at the risk-free rate.
• As S moves, $Δ_t$ changes. Rebalance instantaneously to maintain the hedge ratio. Each rebalancing is self-financing (no cash injections or withdrawals).
• At expiry, the portfolio value exactly equals $H(S_T)$ regardless of the path taken by S

The replication portfolio’s daily P&L has two components. The delta-hedged position eliminates first-order risk. What remains is the second-order term, the gamma P&L:

where

A long gamma position profits from large moves in the underlying (the realised variance exceeds the implied variance priced in by the option’s theta). Theta and gamma are in constant tension, the Black-Scholes equation is precisely the statement that they balance exactly when volatility equals its implied value.

Completeness and the uniqueness of the hedge

In a complete market, every contingent claim can be replicated by a unique self-financing strategy. The martingale representation theorem guarantees that for any square-integrable payoff H, there exists a unique adapted process $φ_t$ such that:

The integrand $φ_t$ is the replicating strategy expressed in terms of the Brownian motion, and it is unique. This is the theoretical guarantee that perfect hedging exists, that there is one and only one correct hedge ratio at every moment in time.

The replication argument links together every piece of the theory: the no-arbitrage condition, the existence of the EMM, Girsanov’s change of measure, and the martingale representation theorem. Together, they form a closed, self-consistent system in which pricing and hedging are two aspects of the same mathematical object.

When arbitrage-free theory breaks down

The arbitrage-free framework assumes a world that does not exist. It posits perfectly liquid markets, frictionless transactions, continuous trading, unlimited borrowing at the risk-free rate, and no counterparty risk. Each of these assumptions is an idealisation. When they fail and in stressed markets, they all fail simultaneously, the elegant mathematics above can become not just inaccurate but dangerously misleading.

Failure mode 1: Funding constraints

The theory assumes you can borrow and lend at the risk-free rate without limit. In practice, banks face balance sheet constraints, regulatory capital requirements and collateral posting obligations. The cost of funding a hedge is not the risk-free rate, it is the firm’s own funding spread (the FVA, or funding valuation adjustment). When funding is constrained, the self-financing condition breaks down: rebalancing the hedge requires cash that may not be available, or may be available only at a significantly higher rate.

Failure mode 2: Liquidity dried up

Perfect replication requires continuous trading in a liquid market. In crises, liquidity can evaporate entirely. Bid-ask spreads widen from basis points to percentages. The act of rebalancing the hedge moves the market against you. In illiquid conditions, the realised cost of dynamic hedging can far exceed the theoretical option premium, destroying the replication argument entirely. The EMM implicitly assumes you can trade at the midpoint, at any size, at any time.

Failure mode 3: Counterparty risk

The theory prices derivatives as if counterparties never default. In reality, an over-the-counter derivative contract is only as good as the counterparty’s ability to pay. Credit Valuation Adjustment (CVA) is the correction to the risk-free derivative price that accounts for the probability that the counterparty defaults before the contract matures. CVA can be large and highly nonlinear, it was a major source of losses in 2008, even for institutions that were otherwise well-hedged.

Failure mode 4: Markets are not frictionless

Transaction costs, taxes, short-selling restrictions, margin requirements, and position limits all drive a wedge between theoretical and realised hedge costs. Rather than a unique arbitrage-free price, the true no-arbitrage condition in the presence of frictions yields an interval, a band of no-arbitrage prices bounded by the cost of super-replication strategies. Inside this band, pricing is indeterminate without additional assumptions about preferences.

The XVA revolution: pricing frictions explicitly

The response of the industry to these failures has been the development of XVA (valuation adjustment) frameworks, which attempt to price the frictions explicitly as add-ons to the risk-neutral value:

CVA = Credit Valuation Adjustment (counterparty default risk)

DVA = Debt Valuation Adjustment (own default risk)

FVA = Funding Valuation Adjustment (funding spread)

ColVA = Collateral Valuation Adjustment (collateral optionality)

KVA = Capital Valuation Adjustment (regulatory capital cost)

The XVA framework is itself contentious, there is no consensus on whether all these adjustments represent true economic costs or accounting artifacts, and they violate the law of one price (different dealers will quote different XVA adjustments for the same trade). But they represent the practitioner community’s acknowledgement that the arbitrage-free world is an idealisation.

The most dangerous application of arbitrage-free theory is to treat the risk-neutral price as if it were the liquidation value of a position. In stressed conditions, the gap between the model price and the price at which a position can actually be unwound can be the difference between solvency and insolvency. Models that assume frictionless markets systematically understate tail risk.