Here is the strangest, deepest idea in mathematical finance: to price an option you do not need to know how likely anything actually is. You need a second set of probabilities — a carefully rigged measure \mathbb{Q} living on the very same outcomes as the real world — chosen so that, after discounting, every traded price becomes a fair game. Same world, different odds: \mathbb{Q} agrees with the real-world measure \mathbb{P} about what can and cannot happen, and about essentially nothing else.

A bookmaker already knows this trick. The odds chalked on the board are not the horses' true chances — they are rigged so that, whichever horse wins, the book breaks even, and every bet is quoted as an expected payout under those rigged odds, no opinion about the race required. The risk-neutral measure is the market's bookmaker's board: find the odds under which discounted prices are martingales, and pricing any claim collapses to taking one expectation.

Discount, then change measure

A bank account growing at the riskless rate r turns 1 today into e^{rt} at time t. To compare a risky price S_t against that benchmark we strip out the guaranteed growth — we discount — giving the discounted price

Under the real-world measure \mathbb{P} a stock typically drifts upward faster than the bank (that excess is its risk premium), so \tilde{S}_t drifts up too — it is not fair. The central trick of arbitrage pricing is to change measure: find a new probability measure \mathbb{Q}, equivalent to \mathbb{P}, under which the discounted price is a fair game.

An equivalent martingale measure (EMM), also called a risk-neutral measure, is a measure \mathbb{Q} \sim \mathbb{P} under which the discounted price is a martingale:

The word equivalent is doing real work: by the Radon–Nikodym relation, \mathbb{Q} and \mathbb{P} must agree on what is possible — we are allowed to reweight the odds, never to conjure or forbid an outcome. Formally: the two measures have exactly the same null sets, \mathbb{P}(A) = 0 \iff \mathbb{Q}(A) = 0. Everything the real world rules out, the pricing world rules out too, and vice versa — but on the events that are possible, their probabilities are free to disagree wildly.

Pricing by discounted expectation, derived line by line

Here is the payoff for all that machinery. Consider a claim that pays H_T at maturity T (a call option pays (S_T - K)^+, say). Let V_t be its arbitrage-free price at time t. We will show

— today's price is just the discounted expected payoff under \mathbb{Q}.

Step 1 — discount the value process. Define the discounted price of the claim itself, \tilde{V}_t = e^{-rt} V_t. The no-arbitrage principle says any tradeable asset must be priced consistently with the stock, so — exactly like \tilde{S}_t — the discounted claim price is a \mathbb{Q}-martingale:

Step 2 — apply the martingale property from 0 to T. Take s = 0 and t = T. Conditioning on \mathcal{F}_0 (which knows nothing but constants) is just an ordinary expectation:

Step 3 — substitute the discount factors. At time 0 there is nothing to discount, so \tilde{V}_0 = e^{0}V_0 = V_0. At maturity the claim is worth its payoff, V_T = H_T, so \tilde{V}_T = e^{-rT} H_T. Plugging both in,

the last equality pulling out the constant discount factor. Price a claim by taking its expected payoff under \mathbb{Q} and discounting. The whole subject of derivative pricing is the project of computing this one expectation for ever more elaborate H_T — the Black–Scholes formula itself is nothing but this expectation, worked out in closed form for a lognormal S_T.

Worked example: a market small enough to solve by hand

Everything above becomes concrete in the tiniest possible market — one step, two outcomes. A stock is worth S_0 = 100 today. Tomorrow it will be worth either S_u = 120 (the real world says this happens with probability p = 0.8) or S_d = 90 (probability 0.2). The bank turns 100 into 105 for certain, so 1 + r = 1.05.

Step 1 — write the martingale condition. An EMM is a new up-probability q (with down-probability 1 - q) under which the discounted stock is fair: today's price equals the discounted \mathbb{Q}-expectation of tomorrow's,

Step 2 — plug in the numbers. Multiplying up, 100 \times 1.05 = 105 must equal 120q + 90(1 - q) = 90 + 30q:

Step 3 — read off the general formula. The same rearrangement with letters, writing u = S_u / S_0 and d = S_d / S_0 for the up and down growth factors, gives the formula every quant knows by heart:

Pause on what just happened. The real-world probability p = 0.8 was stated in the setup — and we never touched it. The rigged odds q are manufactured entirely from prices: u, d and r. Under \mathbb{Q} the stock's expected gross return is \tfrac{1}{2}(120) + \tfrac{1}{2}(90) = 105 — exactly the bank. Risk-neutral indeed.

When is q a genuine probability?

A probability must satisfy 0 < q < 1 (strictly, since equivalence forbids \mathbb{Q} from killing an outcome that \mathbb{P} allows). Reading the formula q = \frac{(1+r) - d}{u - d}, that holds precisely when

i.e. when the certain bank return sits strictly between the stock's worst and best cases. And that is exactly the no-arbitrage condition. Work the boundary cases and watch it fail:

Bank too generous — suppose instead 1 + r = 1.25, so 100 in the bank becomes 125, beating even the stock's best case of 120. The formula gives q = \frac{125 - 90}{30} = \frac{35}{30} \approx 1.17 > 1 — not a probability. The market is telling you why: short the stock, deposit the 100, and tomorrow collect 125 while owing at most 120. A guaranteed profit of at least 5 from nothing.

Bank too stingy — suppose 1 + r = 0.85, below even the stock's worst case of 90. Now q = \frac{85 - 90}{30} = -\frac{1}{6} < 0. The arbitrage runs the other way: borrow 100, buy the stock, and tomorrow hold at least 90 while owing only 85.

So in this two-state world the first fundamental theorem is visible to the naked eye: no arbitrage \iff d < 1 + r < u \iff q \in (0, 1) exists. And the second theorem too: two states, two assets — the equation for q has exactly one solution, so the EMM is unique and the market complete. (Add a third possible price tomorrow without adding an asset and an interval of valid q's appears: many EMMs, an incomplete market, a range of arbitrage-free prices.)

Price a call — and notice what is missing

Back to the honest market (1 + r = 1.05, q = \tfrac{1}{2}). Price a call option with strike K = 100: it pays (S_T - 100)^+, which is 20 in the up state and 0 in the down state. The pricing formula is one line:

Now the shock. The stock goes up 80% of the time — surely that should make the call dearer? The "natural" guess, discounting the real-world expected payoff, gives \frac{0.8 \times 20}{1.05} \approx 15.24. It is wrong, and anyone quoting it gets picked off. The real-world probability p — the drift, the forecast, the analyst's optimism — never appears in the price.

Why: the price is enforced by a hedge, and the hedge never asks about p. Try to manufacture the call's payoff from \Delta shares plus a bank position. Matching both states:

and to match the down state exactly, borrow B today with \tfrac{2}{3}(90) - 1.05B = 0, i.e. B = \frac{60}{1.05} \approx 57.14. Check the up state: \tfrac{2}{3}(120) - 60 = 20. ✓ The portfolio "hold \tfrac{2}{3} of a share, owe 57.14" pays exactly what the call pays, whichever way the coin lands — so today it must cost exactly what the call costs:

The same number, from pure replication — and p appeared in neither equation. Hedging matches the payoff state by state, and with the randomness goes any role for how likely each state was. The risk-neutral q is just the bookkeeping device that turns this replication argument into a one-line expectation.

The Radon–Nikodym weight, made explicit

Changing measure is not magic — it is multiplication. The Radon–Nikodym derivative Z = \frac{d\mathbb{Q}}{d\mathbb{P}} is the per-outcome reweighting factor, and in a two-outcome world you can write it down completely. With p = 0.8 and q = \tfrac{1}{2}:

The up outcome — over-represented in the real world relative to the rigged odds — is dialled down to 62.5\% of its weight; the down outcome is dialled up two-and-a-half-fold. Two sanity checks, both instant:

so \mathbb{Q} is a genuine probability measure (total mass one); and the change-of-measure identity \mathbb{E}_{\mathbb{Q}}[X] = \mathbb{E}_{\mathbb{P}}[Z X] really does make the discounted stock fair:

Equivalence is now visible as a statement about Z: both weights are finite and strictly positive. If the real world had p = 1 (the down move impossible under \mathbb{P} but not under \mathbb{Q}), the weight \frac{1-q}{1-p} would blow up and no equivalent measure could exist — you cannot reweight an impossible event into a possible one. In continuous time the same multiplication is performed path by path: Girsanov's theorem builds the density process Z_t that tilts a Brownian motion's drift away, which is precisely how the figure below is drawn.

Two measures, one path

The same chance outcome \omega tells two stories. Under \mathbb{P} the discounted price \tilde{S}_t carries the stock's excess drift \mu - r and climbs on average; under the equivalent measure \mathbb{Q} that drift is reweighted to zero and \tilde{S}_t merely wanders — a martingale around its start. Step through to see the upward-drifting \mathbb{P} path and the level \mathbb{Q} path drawn from the same Brownian shocks, and Refresh for a fresh \omega.

Notice what the change of measure does not do: the wiggles are identical, because \mathbb{Q} may not invent new paths or delete old ones (equivalence!). All it may re-tilt is the tendency. This is the continuous-time image of the two-outcome weights 0.625 and 2.5 you computed above — a Radon–Nikodym density smoothly dialling each path's probability up or down until the fair-game dashed line is the average.