A share grows at 10% a year on average, but its option is priced as if it grew at the 4% interest rate. Nothing about the share's possible paths has changed: only the probabilities attached to them. This atom shows how to swap one set of probabilities for another, and why the choice of unit for prices decides which set to use.
A ticket pays 10 if a coin lands heads. Under one probability measureProbability measure: a rule giving each set of outcomes a probability. Two measures can share the same outcomes and differ only in the weights., ℙ, the coin is fair. Under another, ℚ, heads comes up 70% of the time. The outcomes are the same; only the weights differ. The ratio of the weights, outcome by outcome, is the likelihood ratioLikelihood ratio Z = dℚ/dℙ: the factor that turns ℙ probabilities into ℚ probabilities, outcome by outcome. Its formal name is the Radon-Nikodym derivative. Z:
Outcome
Payoff X
ℙ
ℚ
Z = ℚ/ℙ
ℙ × Z × X
Heads
10
0.5
0.7
1.4
7
Tails
0
0.5
0.3
0.6
0
E_ℙ[X] = 5 and E_ℚ[X] = 7. The last column adds up to 7: the ℚ-average, computed with ℙ probabilities and the weights Z. And Z itself averages to 1 under ℙ. So you can learn about ℚ without ever tossing a ℚ-coin: toss the fair one 10,000 times (4,911 heads here), multiply each payoff by its Z and average. The answer is 6.875, with a standard error of 0.07, against the true 7.
Everything in this atom is this one move, made on Brownian paths: keep the paths, change the weights.
The rules
The likelihood ratio. Why E_ℚ[X] = E_ℙ[ZX] always works, and what Z must satisfy.
Z exists only if ℚ gives no weight to outcomes that ℙ calls impossible. Measures that agree on what is impossible are equivalentEquivalent measures: two measures that agree on which outcomes are impossible. Only then can each be reached from the other by a positive Z.; then 1/Z takes you back from ℚ to ℙ.
Reweighting has a price. Draws from ℙ multiplied by Z estimate a ℚ-expectation: that is importance samplingImportance sampling: drawing from a convenient measure and multiplying each draw by the likelihood ratio, to estimate an expectation under the measure you care about.. Uneven weights waste draws, because a few heavy draws dominate the average. The effective sample sizeEffective sample size: (ΣZ)²/ΣZ² for N weighted draws, the number of equally weighted draws that would be as informative. About N/E_ℙ[Z²]. measures the loss:
Neff=∑iZi2(∑iZi)2≈EP[Z2]N
For the coin, E_ℙ[Z²] = ½(1.4² + 0.6²) = 1.16, so 10,000 weighted tosses are worth about 8,621 plain ones (the sample gives 8,604). The cost can be worth paying when the new draws land where the payoff lives: worked example 2 shows both sides.
Girsanov. For Brownian motion the likelihood ratio is a ratio of two normal densities, and it moves the drift.
For the Girsanov weight, E_ℙ[Z²] = eθ²T, so N weighted paths are worth about N e−θ²T: a large drift change over a long horizon is expensive to reach by reweighting.
Numeraires. A numeraireNumeraire: a traded asset with a positive price, used as the unit in which other prices are measured. is the unit prices are counted in. For each one there is a measure under which every price divided by the numeraire is a martingaleMartingale: a process whose expected future value, given what is known today, is its value today. It has zero drift.. The money-market accountMoney-market account B(t): £1 invested at t = 0 and rolled over at the short rate, B(t) = exp(∫₀ᵗ r(s) ds). gives the risk-neutral measureRisk-neutral measure ℚ: the measure that goes with the money-market account, under which every price divided by B(t) is a martingale.; a zero-coupon bond gives the T-forward measureT-forward measure ℚᵀ: the measure that goes with the zero-coupon bond P(t, T), under which every price divided by P(t, T), its forward price for date T, is a martingale..
The formal version
Z=dPdQ>0,EP[Z]=1,EQ[X]=EP[ZX]Z(T)=exp(−∫0Tθ(u)dW(u)−21∫0Tθ(u)2du)⇒W(t)=W(t)+∫0tθ(u)du is a Q-Brownian motionN(t)V(t)=EN[N(T)V(T)Ft],dQNdQM=N(T)/N(0)M(T)/M(0)dQdQT=B(T)P(0,T)1,V(0)=EQ[B(T)V(T)]=P(0,T)ET[V(T)]
ℙ and ℚ are equivalent measures on the same outcomes; X is any random amount with a finite ℚ-expectation. W is a ℙ-Brownian motion and θ(t) is known at each time t (constant in this atom); Girsanov needs a mild bound on θ, which a constant meets. V is the price of a traded asset, N and M are numeraires (positive prices), and EN[· | 𝓕t] is the expectation under the measure ℚN given everything known at time t. B(t) is the money-market account with B(0) = 1, P(t, T) the zero-coupon bond paying 1 at T, and ET the expectation under ℚT. Sources: Shreve (2004), volume II, sections 1.6 (change of measure), 5.2.1 (Girsanov's theorem), 9.2 (numéraire) and 9.4 (forward measures); Geman, El Karoui and Rochet (1995) for the change of numeraire.
Try it
2,000 Brownian paths with no drift, drawn once. Move θ and watch the weights, not the paths, change. Then push θ²T past 2 and see what reweighting cannot do cheaply.
Try itReweight Brownian paths to give them a drift
Horizon T
40 of the paths, fainter = less weight Reweighted average of 2,000 paths Target −θt Plain average
Target E_ℚ[W(T)] = −θT−0.5
Reweighted average−0.462 ± 0.032
Plain average (no weights)0.028
Average weight (should be 1)0.985
Spread under ℚ (should be √T)0.987 vs 1
Effective sample size1,564 of 2,000
Theory: N e^(−θ²T)1,558
Z = exp(−θW(T) − ½θ²T) = exp(−0.5·W(T) − 0.125)
No path moves. Paths that end on the side of −θT gain weight, the others lose it, and the weighted average follows the new drift. The spread stays √T: Girsanov changes the drift, not the volatility.
θ²T = 0.25. The weights cost precision: 2,000 weighted paths are worth about 1,558 unweighted ones. Push θ or T up and watch it fall.
Now the numeraire. Three states for year 1, with the money-market account and a 2-year bond in each. Widen the rate move and compare averages under ℚ with averages under the two forward measures.
Try itOne set of states, two numeraires
State
ℚ
B(1)
L(1)
P(1, 2)
ℚ¹
ℚ²
Rates down
0.2500
1.0300
2.00%
0.98039
0.2524
0.2573
Unchanged
0.5000
1.0400
4.00%
0.96154
0.5000
0.4998
Rates up
0.2500
1.0500
6.00%
0.94340
0.2476
0.2429
ℚ¹ uses the 1-year bond as numeraire (weights ℚ/(B(1)P(0, 1))); ℚ² uses the 2-year bond (weights ℚ·P(1, 2)/(B(1)P(0, 2))). Today P(0, 1) = 0.96158 and P(0, 2) = 0.92486.
Forward bond price
P(0, 2)/P(0, 1)0.961805
E¹[P(1, 2)]0.961805
E_ℚ[P(1, 2)]0.961716
Gap under ℚ−0.000089
Forward rate for year 2
P(0, 1)/P(0, 2) − 13.971%
E²[L(1)]3.971%
E_ℚ[L(1)]4.000%
Gap under ℚ2.89bp
Call on P(1, 2), strike 0.96, paid at time 1
E_ℚ[payoff / B(1)]0.005689
P(0, 1) · E¹[payoff]0.005689
P(0, 1) · E_ℚ[payoff] (the shortcut that fails)0.005642
Low rates mean a low bank account and a high bond price together. ℚ¹ leans towards the states where B(1) is small, which is exactly what makes the forward price its average. Under ℚ the discount factor and the payoff move together, so it cannot be pulled out of the expectation.
Worked examples
1. Three outcomes, two measures
Step 1 of 4
1
The set-up
A price ends at 80, 100, 130. Under ℙ the probabilities are 0.2, 0.5, 0.3; under ℚ they are 0.3, 0.5, 0.2.
2. Importance sampling a rare event
Step 1 of 4
1
The target
P(X > 3) for X ~ N(0, 1) is 1 − Φ(3) = 0.00135. With 10,000 plain draws you expect only about 13 hits.
3. Girsanov, checked by simulation
Step 1 of 4
1
The weights
10,000 draws of W(1) under ℙ, each weighted by Z = exp(−0.5W(1) − ½ × 0.5² × 1). The average weight is 1.0053.
4. Changing numeraire in three states
Step 1 of 5
1
The states
Over year 1 rates fall to 2%, stay at 4% or rise to 6%, with ℚ-probabilities 0.25, 0.5, 0.25. £1 rolled overnight earns the year's average rate, so B(1) = 1.03, 1.04, 1.05. A bond paying 1 at time 2 is then worth P(1, 2) = 1/(1 + L(1)) = 0.98039, 0.96154, 0.9434. The numbers are illustrative.
Common mistakes
Each line below is a tempting slip. Spot what is wrong, then check.
Reaction
Check what stuck
Question 1 · objective 1
A coin is fair under ℙ. Under ℚ it lands heads with probability 0.8. What is the likelihood ratio Z on tails?
Question 2 · objective 1
Under ℙ, X is 0, 10 or 20 with probabilities 0.25, 0.5, 0.25. The likelihood ratio on those outcomes is 2, 0.8 and 0.4. What is E_ℚ[X]?
Question 3 · objective 2
You reweight 10,000 Brownian paths with θ = 1 over T = 1 year. About how many equally weighted paths are they worth (the effective sample size)?
Question 4 · objective 3
W is a Brownian motion under ℙ, and ℚ is defined by Z = exp(−θW(T) − ½θ²T). How is W(T) distributed under ℚ?
Question 5 · objective 4
Under the T-forward measure ℚᵀ, which of these is a martingale (for a bond maturing at U > T)?
Glossary
Probability measure
Probability measure: a rule giving each set of outcomes a probability. Two measures can share the same outcomes and differ only in the weights.
Likelihood ratio (Radon-Nikodym derivative)
Likelihood ratio Z = dℚ/dℙ: the factor that turns ℙ probabilities into ℚ probabilities, outcome by outcome. Its formal name is the Radon-Nikodym derivative.
Equivalent measures
Equivalent measures: two measures that agree on which outcomes are impossible. Only then can each be reached from the other by a positive Z.
Importance sampling
Importance sampling: drawing from a convenient measure and multiplying each draw by the likelihood ratio, to estimate an expectation under the measure you care about.
Effective sample size
Effective sample size: (ΣZ)²/ΣZ² for N weighted draws, the number of equally weighted draws that would be as informative. About N/E_ℙ[Z²].
Girsanov's theorem
Girsanov's theorem: reweighting Brownian paths by Z = exp(−θW(T) − ½θ²T) makes W(t) + θt a Brownian motion under the new measure. The drift changes; the volatility does not.
Numeraire
Numeraire: a traded asset with a positive price, used as the unit in which other prices are measured.
Martingale
Martingale: a process whose expected future value, given what is known today, is its value today. It has zero drift.
Money-market account
Money-market account B(t): £1 invested at t = 0 and rolled over at the short rate, B(t) = exp(∫₀ᵗ r(s) ds).
Risk-neutral measure
Risk-neutral measure ℚ: the measure that goes with the money-market account, under which every price divided by B(t) is a martingale.
T-forward measure
T-forward measure ℚᵀ: the measure that goes with the zero-coupon bond P(t, T), under which every price divided by P(t, T), its forward price for date T, is a martingale.
Shreve, Stochastic Calculus for Finance II, section 9.3: the foreign and domestic risk-neutral measures, a change of numeraire between currencies. Advanced.
Glasserman, Monte Carlo Methods in Financial Engineering, chapter 4: importance sampling among the other variance reduction methods. Advanced.
Origins
Steven E. Shreve, Stochastic Calculus for Finance II: Continuous-Time Models, Springer Finance, Springer, 2004 (section 1.6, change of measure; section 5.2.1, Girsanov's theorem; chapter 9, change of numéraire, with sections 9.2 and 9.4).
Tomas Björk, Arbitrage Theory in Continuous Time, 4th edition, Oxford University Press, 2019 (chapter 15, change of numeraire).
Hélyette Geman, Nicole El Karoui and Jean-Charles Rochet, “Changes of numéraire, changes of probability measure and option pricing”, Journal of Applied Probability 32(2), 443–458, 1995.
Paul Glasserman, Monte Carlo Methods in Financial Engineering, Springer, 2003 (chapter 4, variance reduction techniques).