From Theory to the Vasicek Model
A first-principles route from the short rate and stochastic calculus to bond prices under the Vasicek model.
Modelling the yield curve
In our previous post on Belgian government debt (OLOs), we introduced the yield curve, which determines the prices of bonds of different maturities. Equivalently, it determines what one euro paid at a future date \(T\) is worth today. This value is the discount factor; it equals the price \(P(t,T)\) at time \(t\) of a zero-coupon bond paying one euro at \(T\).
The yield curve moves over time, driven by the factors discussed in the previous post, so we need a model of how it may evolve. Modelling the entire curve directly is difficult, since it is a whole function of maturity. Instead, we model a single quantity, the short rate \(r_t\): the interest rate earned over an infinitesimally short period starting at time \(t\). Under the assumption of no arbitrage, the whole curve can then be reconstructed from the short rate.
If the future path of the short rate were known, one euro paid at \(T\) would be worth, at time \(t\), its value discounted at the continuously compounded short rate:
The future path is not known, however, so \(D(t,T)\) is random. The bond price is obtained by averaging the discount factor over all possible future paths, given the information available today:
Two features of this formula require explanation: the expectation is taken under a particular probability measure \(\mathbb Q\), the risk-neutral measure, and it is conditional on \(\mathcal F_t\), the information available at time \(t\). Both are introduced step by step below.
We model the short rate with the Vasicek stochastic differential equation (SDE):
It consists of two building blocks.
- Drift: \(\kappa(\theta-r_t)\,\mathrm dt\). Here \(\mathrm dt\) is an infinitesimal increment of time, \(\kappa>0\) is the speed of mean reversion, \(\theta\) is the long-run level and \(r_t\) is the current short rate. Given the current rate, the drift is known exactly: it pulls the rate towards \(\theta\), downwards when \(r_t>\theta\) and upwards when \(r_t<\theta\). It is not deterministic over time, however, since \(r_t\) is itself random.
- Noise: \(\sigma\,\mathrm dW_t\). Here \(\sigma\) is the volatility and \(\mathrm dW_t\) is an infinitesimal increment of a Brownian motion (Wiener process) \(W\).[¹]
Over a small but finite time step \(\Delta t\), the equation is approximated by
with a new, independent \(\varepsilon\) drawn at every step. This approximation, the Euler–Maruyama scheme, is also how the model is simulated in practice: draw \(\varepsilon\), update the rate, and repeat.
Before we can use this equation to simulate future yield curves, we need to make precise what a scenario is, what a probability is, what it means for a quantity to be random, and how an equation involving \(\mathrm dW_t\) should be read. The following sections build up this framework.
The probability space
In stochastic calculus we work with a probability space \((\Omega,\mathcal F,\mathbb P)\).
\(\Omega\) is the sample space: the set of all possible scenarios. Its elements \(\omega\in\Omega\) are the individual scenarios. In the interest-rate setting, a scenario is a complete future path of the short rate, so \(\Omega\) is typically uncountably infinite, for instance the set of all continuous paths.
\(\mathcal F\) is a collection of subsets of \(\Omega\), that is, \(\mathcal F\subseteq 2^{\Omega}\). Its elements \(A\in\mathcal F\) are called events. An event formalises a statement about scenarios, such as "the short rate exceeds 4% within the next five years", by identifying it with the set of scenarios in which the statement is true. When \(\Omega\) is uncountable, probabilities cannot in general be assigned consistently to every subset (we return to this point below). We therefore restrict attention to a collection of subsets with suitable closure properties, studied in the branch of mathematics called measure theory: a σ-algebra.[²]
Definition 1.1.1 (σ-algebra). \(\mathcal F\) is a σ-algebra if
- \(\emptyset\in\mathcal F\);
- \(A\in\mathcal F\Rightarrow A^{c}\in\mathcal F\);
- \(A_1,A_2,\ldots\in\mathcal F\Rightarrow\bigcup_{n=1}^{\infty}A_n\in\mathcal F\).
These conditions ensure that statements can be combined. If a statement is an event, so is its negation (condition 2). If each of a sequence of statements is an event, so is the statement that at least one of them holds (condition 3). By De Morgan's laws, the statement that all of them hold is then an event as well.
Once the events are fixed, we assign probabilities to them through a probability measure.
Definition 1.1.2 (probability measure). \(\mathbb P:\mathcal F\to[0,1]\) is a probability measure if
- \(\mathbb P(\Omega)=1\);
- for pairwise disjoint \(A_1,A_2,\ldots\in\mathcal F\): \(\mathbb P\big(\bigcup_n A_n\big)=\sum_n\mathbb P(A_n)\) (countable additivity).
Consequences: \(\mathbb P(\emptyset)=0\); \(\mathbb P(A^{c})=1-\mathbb P(A)\); \(A\subseteq B\Rightarrow\mathbb P(A)\le\mathbb P(B)\) (monotonicity); and \(\mathbb P\big(\bigcup_n A_n\big)=\lim_n\mathbb P(A_n)\) for \(A_1\subseteq A_2\subseteq\cdots\) (continuity from below).
The triple \((\Omega,\mathcal F,\mathbb P)\) is a probability space:
Application. The same pair \((\Omega,\mathcal F)\) can carry several probability measures. Two of them matter here.
- \(\mathbb P\), the real-world measure, describes how likely the various interest-rate scenarios are believed to be. It is a model, calibrated to historical data and expert judgement, and it is used for projections. The scenarios in the reinforcement-learning environment are generated under \(\mathbb P\).
- \(\mathbb Q\), the risk-neutral measure, is used for pricing. Bond prices, and hence the OLO yields, are \(\mathbb Q\)-expectations, as in the bond-price formula above.
- Random variables, such as the short rate \(r_t\), are the same functions under both measures; only their distributions differ. The precise relation between \(\mathbb P\) and \(\mathbb Q\) is the subject of the section on change of measure.
Random variables
Definition (random variable). Let \((\Omega,\mathcal F,\mathbb P)\) be a probability space. A random variable is a function \(X:\Omega\to\mathbb R\) such that
This property is called \(\mathcal F/\mathcal B(\mathbb R)\)-measurability. Note that it involves only \(\mathcal F\), not \(\mathbb P\).
Interpretation. A random variable assigns a real number to each scenario. \(X(\omega)\) is the value that the quantity of interest, for instance the annual return on a pension reserve, takes in scenario \(\omega\). Distinct scenarios may produce the same value, so \(X\) need not be invertible.
Not every function qualifies. Each set \(B\) of values represents a question about \(X\), for instance "does the return fall below the guaranteed rate?". The set \(\{X\in B\}\) collects precisely the scenarios in which the answer is affirmative. Since \(\mathbb P\) assigns probabilities only to sets in \(\mathcal F\), the requirement \(\{X\in B\}\in\mathcal F\) ensures that every such question has a well-defined probability.
The Borel σ-algebra. The questions \(B\) are drawn from \(\mathcal B(\mathbb R)\), the Borel σ-algebra: the smallest σ-algebra on \(\mathbb R\) containing all intervals. It contains every interval and is closed under complements, countable unions and countable intersections. That is, applying any of these operations to Borel sets, including sets that were themselves obtained through earlier such operations, always yields another Borel set. Here countable means that the sets involved can be listed as a sequence \(B_1,B_2,B_3,\dots\). The Borel σ-algebra does not contain every subset of \(\mathbb R\).
Why the Borel σ-algebra. The collection of admissible sets \(B\) determines which questions about \(X\) are required to have a probability. Its choice involves a trade-off. If the collection is too small, natural questions cannot be asked. If it is too large, probabilities can no longer be assigned to all of them consistently. The Borel σ-algebra resolves this trade-off.
Why not fewer sets. The most basic question about a quantity is whether it lies at or below a given threshold \(x\), that is, whether \(X\le x\). Answering this for every threshold yields the distribution function \(F_X(x)=\mathbb P(X\le x)\). Combining threshold questions produces further questions. For example, "is the return above 0% but at most 2.5%?" corresponds to the set \((-\infty,2.5\%]\) with \((-\infty,0\%]\) removed. Some natural questions require infinitely many threshold questions. "Is the return strictly below 2.5%?" cannot be expressed through finitely many questions of the form \(X\le x\). It is answered affirmatively exactly when the return lies at or below \(2.5\%-\tfrac1n\%\) for at least one \(n=1,2,3,\dots\), that is, at or below 1.5%, or 2%, or 2.17%, and so on, approaching 2.5% from below without reaching it. This is a countable union:
Similarly, "is the return exactly 2.5%?" requires a countable intersection. Closure under countable operations ensures that all such questions are available, and the Borel σ-algebra is the smallest collection that contains the threshold sets and has this property.
Why not more sets. Closure is restricted to countable combinations deliberately. Every subset of \(\mathbb R\) is a union of single points, so permitting arbitrary unions would admit every subset. That is too demanding. It is known that \(\mathbb R\) contains highly irregular subsets to which no sensible notion of length can be assigned; their existence can be proven, but they cannot be described explicitly and have no practical interpretation. Requiring a probability for every such set would rule out even the simplest continuous distributions, such as a return drawn uniformly between 0% and 10%. Restricting to Borel sets excludes these pathological cases while retaining every set that arises in practice.[³]
Practical consequence. Because every Borel set is obtained from threshold sets through countable combinations, it suffices to verify that \(\{X\le x\}\in\mathcal F\) for every \(x\in\mathbb R\). Measurability for all Borel sets then follows automatically.
How \(X\) links the probability space to the value space
The random variable plays three roles:
- Outcomes are mapped forward, element by element: $$\omega\ \longmapsto\ X(\omega).$$
- Questions are pulled back, set by set: $\(B\ \longmapsto\ \{X\in B\}=X^{-1}(B)\in\mathcal F.$$ Here\)X^{-1}(B)\(denotes the *preimage* of\)B\(, the set of all scenarios whose value lies in\)B$. It is defined for every function, invertible or not, and should not be confused with an inverse function.
- Probability is transported forward: $$\mu_X(B):=\mathbb P{X\in B},\qquad B\in\mathcal B(\mathbb R).$$
The first two roles are linked by
The distribution of \(X\)
\(\mu_X\) is a probability measure on \((\mathbb R,\mathcal B(\mathbb R))\), called the distribution or law of \(X\). This holds because \(\mu_X(\mathbb R)=\mathbb P(\Omega)=1\), and because preimages preserve complements, countable unions and disjointness, so the axioms satisfied by \(\mathbb P\) carry over to \(\mu_X\).
Moreover, \(\mu_X\) is uniquely determined by the cumulative distribution function
Knowing \(F_X\) therefore suffices to compute the probability of every Borel set.
The distribution is specific to the variable. Two random variables on the same probability space generally induce different distributions on the same value space. A Borel set therefore carries no probability of its own; it acquires one only through a random variable.
Expectation
Let \(X\) be a random variable on \((\Omega,\mathcal F,\mathbb P)\). If \(X\ge 0\), or if \(\mathbb E\lvert X\rvert<\infty\), the expectation of \(X\) is defined as
Intuitively, \(\mathbb E X\) is a probability-weighted average of the values \(X(\omega)\) over all scenarios \(\omega\). For \(X\ge 0\) the integral is always defined, although it may equal \(+\infty\). For general \(X\) the condition \(\mathbb E\lvert X\rvert<\infty\) is required; otherwise large positive and large negative values may fail to balance, and the average is undefined.
Theorem (change of variables). Let \(X\) be a random variable on \((\Omega,\mathcal F,\mathbb P)\) with distribution \(\mu_X\), and let \(g:\mathbb R\to\mathbb R\) be Borel measurable. Then
where both sides may equal \(+\infty\). If this quantity is finite, then
The first statement is used to verify that \(\mathbb E\,g(X)\) exists; the second computes it. This result is also known as the law of the unconscious statistician.
Remark. The theorem replaces an integral over the abstract scenario space \(\Omega\) by an integral over the real line. Its significance is that the expectation of any function of \(X\) depends only on the distribution \(\mu_X\), not on the underlying probability space. Two models that assign the same distribution to, for instance, the return on a pension reserve yield the same expected value for every function of that return, however differently they describe the scenarios. In particular, once \(\mu_X\) is known or can be sampled from, \(\Omega\) need not be specified explicitly.
The right-hand side is still a Lebesgue integral, but with respect to \(\mu_X\) rather than Lebesgue measure (ordinary length). An explicit computation requires a further reduction, which depends on the form of \(\mu_X\):
-
Distributions with a density. If there exists a function \(f_X\ge 0\) such that \(\mu_X(B)=\int_B f_X(x)\,\mathrm dx\) for all \(B\in\mathcal B(\mathbb R)\), then \(X\) is called absolutely continuous with density \(f_X\), and
\[\mathbb E\,g(X)=\int_{-\infty}^{\infty}g(x)\,f_X(x)\,\mathrm dx.\]The density plays the role of a weight attached to each value \(x\). Having a continuous distribution function \(F_X\) does not by itself guarantee that a density exists.[⁴]
-
Discrete distributions. If \(X\) takes only the distinct values \(x_1,x_2,\dots\), with probabilities \(p_i=\mathbb P\{X=x_i\}\) satisfying \(\sum_i p_i=1\), then
\[\mathbb E\,g(X)=\sum_i g(x_i)\,p_i,\]provided \(\sum_i \lvert g(x_i)\rvert\,p_i<\infty\).
-
Mixed distributions. Many quantities of interest are neither purely discrete nor absolutely continuous. A benefit of the form \(\max(A_T,G_T)\), where \(A_T\) is the accumulated account value and \(G_T\) the guaranteed amount, equals \(G_T\) with positive probability and varies continuously above it. In such cases the expectation combines both forms: a sum over the values carrying positive probability, plus an integral against a density over the remaining values.
Change of measure
Motivation. The real-world measure \(\mathbb P\) and the risk-neutral measure \(\mathbb Q\) are defined on the same measurable space \((\Omega,\mathcal F)\). They describe the same scenarios and the same events; they differ only in the probability assigned to each event. A change of measure keeps the scenarios fixed and changes their weights.
Example. Consider two interest-rate scenarios: \(\omega_1\) (rates rise) and \(\omega_2\) (rates fall). Suppose \(\mathbb P(\omega_1)=\mathbb P(\omega_2)=0.5\), while \(\mathbb Q(\omega_1)=0.6\) and \(\mathbb Q(\omega_2)=0.4\). Define the weight \(Z(\omega):=\mathbb Q(\omega)/\mathbb P(\omega)\), so that \(Z(\omega_1)=1.2\) and \(Z(\omega_2)=0.8\). For any random variable \(Y\),
A \(\mathbb Q\)-expectation is thus a \(\mathbb P\)-expectation in which each scenario is reweighted by \(Z\). The weights are positive and average to one under \(\mathbb P\): \(\mathbb E^{\mathbb P}[Z]=0.5\cdot1.2+0.5\cdot0.8=1\).
Definition (equivalent measures). Two probability measures \(\mathbb P\) and \(\mathbb Q\) on \((\Omega,\mathcal F)\) are equivalent, written \(\mathbb Q\sim\mathbb P\), if they assign probability zero to the same events:
Equivalent measures agree on which events are possible; they may disagree only on how likely the possible events are. This is the natural requirement for valuation: a scenario that cannot occur in reality should not affect prices, and conversely.
Theorem (Radon–Nikodym). If \(\mathbb Q\sim\mathbb P\), there exists a random variable \(Z>0\) with \(\mathbb E^{\mathbb P}[Z]=1\) such that
for every random variable \(Y\) for which either side is defined. \(Z\) is called the Radon–Nikodym derivative and is written \(Z=\mathrm d\mathbb Q/\mathrm d\mathbb P\). It generalises the scenario weights of the example to settings in which individual scenarios have probability zero, so the ratio \(\mathbb Q(\omega)/\mathbb P(\omega)\) is no longer defined.
So far the change of measure is static: it reweights scenarios as a whole. How it acts on processes evolving over time, and in particular what it does to the dynamics of interest rates, is treated after stochastic processes and stochastic integrals have been introduced.
Information over time
Stochastic processes. A stochastic process is a collection of random variables \(X=(X_t)_{t\in[0,T]}\) on a common probability space, one for each time \(t\). For a fixed scenario \(\omega\), the map \(t\mapsto X_t(\omega)\) is the path of the process in that scenario: for instance, the evolution of the short rate over the projection horizon.
Definition (filtration). A filtration is a family \((\mathcal F_t)_{t\in[0,T]}\) of σ-algebras with \(\mathcal F_s\subseteq\mathcal F_t\subseteq\mathcal F\) for \(s\le t\). The tuple \((\Omega,\mathcal F,(\mathcal F_t),\mathbb P)\) is called a filtered probability space.
\(\mathcal F_t\) represents the information available at time \(t\): the collection of events whose occurrence can be decided by then. The inclusion \(\mathcal F_s\subseteq\mathcal F_t\) expresses that information is never lost.
A process \(X\) is adapted to \((\mathcal F_t)\) if \(X_t\) is \(\mathcal F_t\)-measurable for every \(t\). In words, the value of \(X_t\) is known at time \(t\). Decisions in a pension plan, such as a contribution rate set at time \(t\), must be adapted: they may depend on the past and present, but not on the future.
Definition (conditional expectation). Let \(Y\) be a random variable with \(\mathbb E\lvert Y\rvert<\infty\), and let \(\mathcal G\subseteq\mathcal F\) be a σ-algebra. The conditional expectation \(\mathbb E[Y\mid\mathcal G]\) is the random variable that
- is \(\mathcal G\)-measurable, and
- satisfies \(\mathbb E\big[\mathbb E[Y\mid\mathcal G]\,\mathbf 1_A\big]=\mathbb E[Y\,\mathbf 1_A]\) for every \(A\in\mathcal G\).
Such a random variable exists and is unique up to events of probability zero.
Interpretation. \(\mathbb E[Y\mid\mathcal F_t]\) is the best forecast of \(Y\) given the information available at time \(t\). Condition 1 states that the forecast uses only that information. Condition 2 states that it is correct on average over every event observable at time \(t\). Unlike an ordinary expectation, it is itself a random variable, since the forecast depends on which scenario has unfolded up to time \(t\). This is why the bond-price formula at the start is a conditional expectation: the price at time \(t\) depends on the short rate observed at \(t\).
Properties. For integrable \(Y\), \(X\) and σ-algebras \(\mathcal H\subseteq\mathcal G\):
- Tower property: \(\mathbb E\big[\mathbb E[Y\mid\mathcal G]\mid\mathcal H\big]=\mathbb E[Y\mid\mathcal H]\). Forecasting today a forecast that will be made tomorrow gives today's forecast. In particular \(\mathbb E\big[\mathbb E[Y\mid\mathcal G]\big]=\mathbb E[Y]\).
- Taking out what is known: if \(X\) is \(\mathcal G\)-measurable and \(XY\) is integrable, then \(\mathbb E[XY\mid\mathcal G]=X\,\mathbb E[Y\mid\mathcal G]\).
- Independence: if \(Y\) is independent of \(\mathcal G\), then \(\mathbb E[Y\mid\mathcal G]=\mathbb E[Y]\). Information unrelated to \(Y\) does not improve the forecast.
Definition (martingale). An adapted process \(M\) with \(\mathbb E\lvert M_t\rvert<\infty\) for all \(t\) is a martingale if
The best forecast of any future value is the current value: the process has no predictable tendency to rise or fall.
Under the risk-neutral measure, the price of every traded asset divided by the bank account \(B(t)=\exp\big(\int_0^t r_s\,\mathrm ds\big)\) is a martingale. For a zero-coupon bond this means that \(P(t,T)/B(t)\) is a \(\mathbb Q\)-martingale. Since \(P(T,T)=1\), this property is equivalent to the bond-price formula at the start of this post. The bond price itself is not a martingale: it rises towards its face value as maturity approaches.
Brownian motion
Interest rates and asset returns are subject to a continuous stream of small, unpredictable shocks. Brownian motion is the basic mathematical model for the cumulative effect of such shocks, and it is the source of randomness in the Vasicek model.
Definition (Brownian motion). A process \(W=(W_t)_{t\in[0,T]}\) on a filtered probability space \((\Omega,\mathcal F,(\mathcal F_t),\mathbb P)\) is a Brownian motion if
- \(W_0=0\);
- \(W\) is adapted to \((\mathcal F_t)\);
- for \(0\le s
, the increment \(W_t-W_s\) is independent of \(\mathcal F_s\): future shocks cannot be predicted from the past; - for \(0\le s
, the increment \(W_t-W_s\) is normally distributed with mean \(0\) and variance \(t-s\); - the paths \(t\mapsto W_t(\omega)\) are continuous for every \(\omega\).
Interpretation. Each scenario \(\omega\) now corresponds to an entire path. Conditions 3 and 4 state that the change over any interval is unpredictable, has no systematic direction, and has a variance proportional to the length of the interval. A useful mental picture is a random walk taking very many very small steps. By the central limit theorem, the sum of many small independent shocks is approximately normal, which motivates condition 4. Condition 4 is also why the noise term in the discretised Vasicek equation is \(\sigma\sqrt{\Delta t}\,\varepsilon\): the increment \(W_{t+\Delta t}-W_t\) has the same distribution as \(\sqrt{\Delta t}\,\varepsilon\) with \(\varepsilon\sim\mathcal N(0,1)\).
Properties.
- Mean and variance: \(\mathbb E[W_t]=0\) and \(\operatorname{Var}(W_t)=t\). Uncertainty grows with the horizon, with the standard deviation growing as \(\sqrt t\).
- Martingale property: \(W\) is a martingale. Indeed, for \(s
, \(\mathbb E[W_t\mid\mathcal F_s]=W_s+\mathbb E[W_t-W_s\mid\mathcal F_s]=W_s+\mathbb E[W_t-W_s]=W_s\), using that \(W_s\) is known at time \(s\) and that the increment is independent of \(\mathcal F_s\) with mean zero. - Irregular paths: although continuous, the paths are nowhere differentiable. Consequently an integral \(\int H_t\,\mathrm dW_t\) cannot be defined path by path as an ordinary integral, and a separate construction is required.
- Quadratic variation: if \([0,t]\) is divided into \(n\) subintervals of equal length, then $$\sum_{k=1}^{n}\big(W_{kt/n}-W_{(k-1)t/n}\big)^2\ \longrightarrow\ t\qquad\text{as }n\to\infty.$$ The squared fluctuations of a single path therefore reveal the elapsed time and, once the process is scaled by a volatility, that volatility as well. This property is used later to explain why a change of measure can alter the drift of a process but never its volatility.[⁵]
The stochastic integral
Construction. Let \(H\) be an adapted process, for instance the amount invested in a risky asset at each time. The gain from holding \(H_t\) units of exposure while \(W\) moves should be \(\int_0^T H_t\,\mathrm dW_t\). For a process that is constant on the intervals of a partition \(0=t_0
The integrand is evaluated at the left endpoint of each interval: the position is chosen with the information available at \(t_k\), before the subsequent shock is observed. This reflects the requirement that decisions cannot anticipate the future. For general adapted \(H\), the integral is defined as the limit of such sums as the partition becomes finer. The result is the Itô integral.
Theorem (properties of the Itô integral). Let \(H\) be adapted with \(\mathbb E\big[\int_0^T H_t^2\,\mathrm dt\big]<\infty\), and let \(I_t:=\int_0^t H_s\,\mathrm dW_s\). Then
- \(\mathbb E[I_t]=0\);
- Itô isometry: \(\mathbb E\big[I_t^2\big]=\mathbb E\big[\int_0^t H_s^2\,\mathrm ds\big]\);
- \((I_t)\) is a martingale.
Remark. The integrability condition is essential. If only \(\int_0^T H_t^2\,\mathrm dt<\infty\) holds scenario by scenario, the integral can still be defined, but it is in general only a local martingale: a process that behaves as a martingale up to suitable random times, but whose expectation may change over time. Not every stochastic integral is therefore a martingale. This distinction becomes important for the change of measure below.
Itô processes and SDEs. An Itô process is a process of the form
with adapted \(\mu\) and \(\sigma\). The first integral is an ordinary integral describing the predictable tendency, or drift. The second is an Itô integral describing the random fluctuations, scaled by the volatility \(\sigma\). This is abbreviated as
The differential notation is shorthand for the integral equation; it does not assert that \(X\) is differentiable. When \(\mu_t=\mu(t,X_t)\) and \(\sigma_t=\sigma(t,X_t)\) depend on the current state, the equation is a stochastic differential equation (SDE). The Vasicek equation from the start of this post is of this form, with \(\mu(t,r)=\kappa(\theta-r)\) and constant \(\sigma\). It is now clear how it should be read:
Theorem (Itô's formula). Let \(X\) be an Itô process as above and let \(f(t,x)\) be once continuously differentiable in \(t\) and twice in \(x\). Then
with the derivatives evaluated at \((t,X_t)\).
Compared with the ordinary chain rule, the additional term \(\tfrac12\partial_{xx}f\,\sigma_t^2\,\mathrm dt\) arises from the quadratic variation of Brownian motion. Over a short interval of length \(\Delta t\), the squared increment \((\Delta W)^2\) is of order \(\Delta t\) rather than negligible, so second-order terms in a Taylor expansion contribute. This is often summarised by the heuristic rule \((\mathrm dW_t)^2=\mathrm dt\).
Change of measure for processes
Density process. Let \(\mathbb Q\sim\mathbb P\) with Radon–Nikodym derivative \(Z=\mathrm d\mathbb Q/\mathrm d\mathbb P\), as introduced earlier. The weight associated with the information available at time \(t\) is
By the tower property, \((Z_t)\) is a \(\mathbb P\)-martingale with \(Z_T=Z\); if \(\mathcal F_0\) contains no information beyond events of probability zero or one, then also \(Z_0=1\). It is called the density process. In the two-scenario example, \(Z\) reweighted entire scenarios. \(Z_t\) refines this to the information available at time \(t\): it is the reweighting of what has been observed so far.
Theorem (Girsanov). Let \(W\) be a Brownian motion under \(\mathbb P\), and let \(\lambda\) be an adapted process. Define
If \((Z_t)\) is a martingale on \([0,T]\), then \(\mathrm d\mathbb Q:=Z_T\,\mathrm d\mathbb P\), that is, \(\mathbb Q(A):=\mathbb E^{\mathbb P}[Z_T\mathbf 1_A]\), defines a probability measure equivalent to \(\mathbb P\). Under \(\mathbb Q\), the process
is a Brownian motion.[⁶]
When is \(Z\) a martingale? By Itô's formula, \(Z\) satisfies \(\mathrm dZ_t=-\lambda_t Z_t\,\mathrm dW_t\), so it is a stochastic integral. As noted in the previous section, stochastic integrals are in general only local martingales. Since \(Z>0\), it can be shown that \(\mathbb E^{\mathbb P}[Z_T]\le1\), with equality precisely when \(Z\) is a true martingale on \([0,T]\). If the inequality is strict, \(\mathbb Q(\Omega)<1\), and \(\mathbb Q\) is not a probability measure. The martingale condition in the theorem therefore cannot be omitted. A sufficient condition is Novikov's criterion,
which holds in particular when \(\lambda\) is bounded.
Effect on the Vasicek model. Consider the Vasicek model under \(\mathbb P\):
With a constant market price of risk \(\lambda\), substituting \(\mathrm dW_t=\mathrm dW^{\mathbb Q}_t-\lambda\,\mathrm dt\) gives
The model keeps the same form under both measures. Only the long-run level towards which rates revert changes; the volatility \(\sigma\) is unaffected.
This holds generally. By the quadratic variation property of Brownian motion, the volatility of a process can be recovered from the fluctuations of a single path. Since equivalent measures agree on which paths are possible, they must agree on the volatility. Only the drift, which reflects how likely the various paths are, can differ. The market price of risk \(\lambda\) quantifies the compensation investors require for bearing interest-rate risk, and it is the parameter that connects the real-world projection model under \(\mathbb P\) to the risk-neutral valuation model under \(\mathbb Q\).
Back to the yield curve
We can now return to the bond-price formula from the start:
Under the Vasicek model this expectation can be computed in closed form. With \(\tau:=T-t\) the remaining maturity,
The corresponding continuously compounded yield is
Two observations follow. First, the entire yield curve at time \(t\) is determined by a single number, the current short rate \(r_t\): this is what it means to reconstruct the curve from the short rate. Second, the formula uses the risk-neutral parameter \(\theta^{\mathbb Q}\), not the real-world \(\theta\), because prices are \(\mathbb Q\)-expectations.
This gives a natural way to simulate future yield curves:
- Simulate paths of the short rate under \(\mathbb P\), using the discretised Vasicek equation with the real-world parameters \(\kappa\), \(\theta\) and \(\sigma\).
- At each future date along each path, evaluate the bond-price formula at the simulated \(r_t\), using the risk-neutral level \(\theta^{\mathbb Q}\).
The result is a collection of realistic yield-curve scenarios, each of which is internally consistent with no-arbitrage pricing.
The model has known limitations. Because a single random factor drives the entire curve, yields of all maturities move in perfect correlation, so the model cannot capture changes in the slope or curvature of the curve that are independent of its level. Moreover, the short rate is normally distributed and can therefore become negative. Depending on the period under study, the latter is a weakness or, given the negative euro-area yields observed in recent years, a realistic feature.
Further reading
If you want to explore these concepts in more depth, the following books are recommended:
- D. Williams, Probability with Martingales, Cambridge University Press — a concise, rigorous introduction to measure-theoretic probability, conditional expectation and martingales.
- S. Shreve, Stochastic Calculus for Finance II: Continuous-Time Models, Springer — Brownian motion, Itô calculus, Girsanov's theorem and risk-neutral pricing, with a financial focus.
- T. Björk, Arbitrage Theory in Continuous Time, Oxford University Press — an accessible treatment of arbitrage pricing, including short-rate models.
- D. Brigo and F. Mercurio, Interest Rate Models — Theory and Practice, Springer — the standard reference on short-rate models, including Vasicek and its extensions.
Footnotes
[¹]: Brownian motion is continuous but nowhere differentiable, so \(\mathrm dW_t\) is not the derivative of a smooth function multiplied by \(\mathrm dt\). This is why stochastic calculus requires its own integral (see the section on the stochastic integral). The two terms of the SDE also scale differently with the time step: the drift scales as \(\Delta t\), while the noise scales as \(\sqrt{\Delta t}\). For \(\Delta t=0.01\), the drift scale is \(0.01\) and the noise scale is \(\sqrt{0.01}=0.1\), ten times larger, and the ratio grows as the step shrinks. On very short time scales, a path is therefore dominated by noise and the drift is barely visible. Over longer horizons the picture reverses. For a process with constant drift, the drift accumulates linearly in time while the noise accumulates only as \(\sqrt t\), so the drift eventually dominates. In the Vasicek model, mean reversion keeps the noise from accumulating indefinitely: the variance of \(r_t\) converges to \(\sigma^2/(2\kappa)\), and the rate fluctuates around \(\theta\) instead of drifting away.
[²]: An accessible introduction to measure theory can be found here.
[³]: The best-known example of such an irregular set is the Vitali set; see also this video.
[⁴]: The standard counterexample is the Cantor distribution, which places all its probability on the Cantor set. Its distribution function is continuous, but the Cantor set has length zero, so no density can exist.
[⁵]: The convergence holds in mean square, and almost surely along suitable sequences of refining partitions.
[⁶]: Sign conventions differ across the literature; some authors define \(W^{\mathbb Q}_t:=W_t-\int_0^t\lambda_s\,\mathrm ds\), which reverses the sign of \(\lambda\) throughout.