The HJM Framework

Interest Rate Modeling




Tutorial Contents

  1. Fixed Income Notation
  2. Fixed Income Probability Measures
  3. Multi-Currency Markets
  4. The HJM Framework
  5. Basic Fixed Income Instruments
  6. Caps, Floors and Swaptions
  7. CMS Swaps, Caps and Floors
  8. Bermudan Swaptions
  9. Structured Notes and Exotic Swaps
  10. Callable Libor Exotics and TARNs
  11. Volatility Derivatives
  12. Yield Curves and Curve Construction
  13. Managing Yield Curve Risk
  14. Vanilla Models With Local Volatility
  15. Vanilla Models With Stochastic Volatility
  16. Introduction to Term Structure Models
  17. One-Factor Short Rate Models I
  18. One-Factor Short Rate Models II
  19. Multi-Factor Short Rate Models
  20. The Libor Market Model I
  21. The Libor Market Model II
  22. Risk Management and Sensitivity Computations
  23. P&L Attribution
  24. Payoff Smoothing
  25. Vegas in the Libor Market Model


In this section, we assign dynamics to the major quantities we have defined thus far. Specifically, we look at the ZCB prices P(t,T)P(t,T) and the instantaneous forward rate f(t,T)f(t,T). By ensuring that the dynamics of these quantities do not admit arbitrage, we find that, when these values are driven by a finite number of Brownian Motions, the dynamics of our various quantities must satisfy certain conditions. This is the crux of the Heath-Jarrow-Morton (HJM) framework. The HJM framework prescribes the way in which the drift terms of our dynamics must relate to the diffusion terms in order to satisfy no-arbitrage.

Bond Price Dynamics

Within the HJM framework, we focus on modeling an entire continuum of TT-indexed bond prices P(t,T)P(t,T) for t[0,T]t \in [0,T], starting from the known initial conditions P(0,T)P(0,T). We consider ZCB prices over a finite time-horizon, i.e. T[0,T]T \in [0, \mathcal{T}], T<\mathcal{T} < \infty, and assume that all models are generated by a dd-dimensional Brownian motion. Let W(t)W(t) represent a dd-dimensional Brownian motion under the risk-neutral measure Q\mathbb{Q}, with numeraire β(t)\beta(t), the continuously compounded money market account. Define the discounted ZCB prices P~(t,T)=P(t,T)β(t)\tilde{P}(t,T) = \tfrac{P(t,T)}{\beta(t)}. In the absence of arbitrage P~(t,T)\tilde{P}(t,T) is a martingale under Q\mathbb{Q}, and the martingale representation theorem then implies that

dP~(t,T)=P~(t,T)σP(t,T)dW(t), \text{d}\tilde{P}(t,T) = -\tilde{P}(t,T)\sigma_P(t,T)^\intercal\text{d}W(t),

where σP(t,T)\sigma_P(t,T) is a dd-dimensional stochastic process adapted to the filtration generated by WW. We assume that σP(t,T)\sigma_P(t,T) is regular enough for P~(t,T)\tilde{P}(t,T) to be a square-integrable martingale, and since we have the boundary conditions P(T,T)=1P(T,T) = 1 for T[0,T]T \in [0,\mathcal{T}], we impose the extra pull to par condition

σP(T,T)=0. \sigma_P(T,T) = 0.

Considering the SDE for β(t)\beta(t), using Ito's lemma shows that

dP(t,T)P(t,T)=d(β(t)P~(t,T))P(t,T)=P~(t,T)dβ(t)+β(t)dP~(t,T)P(t,T)=dβ(t)β(t)+dP~(t,T)P~(t,T)=r(t)dtσP(t,T)dW(t), \begin{aligned} \frac{\text{d} P(t,T)}{P(t,T)} & = \frac{\text{d} \left(\beta(t)\tilde{P}(t,T) \right)}{P(t,T)} \\ & = \frac{\tilde{P}(t,T)\text{d}\beta(t) + \beta(t)\text{d}\tilde{P}(t,T)}{P(t,T)} \\ & = \frac{\text{d}\beta(t)}{\beta(t)} + \frac{\text{d}\tilde{P}(t,T)}{\tilde{P}(t,T)} \\ & = r(t)\text{d}t -\sigma_P(t,T)^\intercal \text{d}W(t), \end{aligned}

where r(t)r(t) is the short rate. Equation

dP(t,T)P(t,T)=r(t)dtσP(t,T)dW(t) \frac{\text{d} P(t,T)}{P(t,T)} = r(t)\text{d}t -\sigma_P(t,T)^\intercal \text{d}W(t)

defines the class of dd-dimensional HJM models.

Another application of Ito's lemma shows that forward bond prices P(t,T,T+τ)=P(t,T+τ)P(t,T)P(t,T,T+\tau) = \tfrac{P(t,T+\tau)}{P(t,T)} must satisfy

dP(t,T,T+τ)=d(P(t,T+τ)P(t,T))=dP(t,T+τ)P(t,T)P(t,T+τ)P(t,T)2dP(t,T)dP(t,T),P(t,T+τ)P(t,T)2+P(t,T+τ)P(t,T)3dP(t,T)=P(t,T+τ)P(t,T)(dP(t,T+τ)P(t,T+τ)dP(t,T)P(t,T)dP(t,T),P(t,T+τ)P(t,T)P(t,T+τ)+dP(t,T)P(t,T)2)=P(t,T,T+τ)((σP(t,T+τ)σP(t,T))dW(t)(σP(t,T+τ)σP(t,T))σP(t,T)dt). \begin{aligned} \text{d}P(t,T,T+\tau) & = \text{d}\left(\frac{P(t,T+\tau)}{P(t,T)}\right) \\ & = \frac{\text{d}P(t,T+\tau)}{P(t,T)} - \frac{P(t,T+\tau)}{P(t,T)^2}\text{d}P(t,T) - \frac{\text{d}\left\langle P(t,T), P(t,T+\tau) \right\rangle}{P(t,T)^2} + \frac{P(t,T+\tau)}{P(t,T)^3}\text{d}\left\langle P(t,T)\right\rangle \\ & = \frac{P(t,T+\tau)}{P(t,T)}\cdot \left(\frac{\text{d}P(t,T+\tau)}{P(t,T+\tau)} -\frac{\text{d}P(t,T)}{P(t,T)} -\frac{\text{d}\left\langle P(t,T), P(t,T+\tau) \right\rangle}{P(t,T)P(t,T+\tau)} + \frac{\text{d}\left\langle P(t,T)\right\rangle}{P(t,T)^2}\right) \\ & = P(t,T,T+\tau) \cdot \bigg( -(\sigma_P(t,T+\tau) - \sigma_P(t,T))^\intercal \text{d}W(t) - \left( \sigma_P(t,T+\tau) - \sigma_P(t,T) \right)^\intercal \sigma_P(t,T) \text{d}t \bigg). \end{aligned}

Hence, under the risk-neutral measure, forward bond prices satisfy the SDE

dP(t,T,T+τ)P(t,T,T+τ)=(σP(t,T+τ)σP(t,T))(σP(t,T)dt+dW(t)).\begin{aligned} \frac{\text{d}P(t,T,T+\tau)}{P(t,T,T+\tau)} & = -(\sigma_P(t,T+\tau) - \sigma_P(t,T))^\intercal (\sigma_P(t,T)\text{d}t + \text{d}W(t)). \end{aligned}

Since, under the TT-forward measure QT\mathbb{Q}^T P(t,T,T+τ)P(t,T,T+\tau) is a martingale, it must be driftless under the QT\mathbb{Q}^T measure, which immediately shows that the QT\mathbb{Q}^T-Brownian motion is related to W(t)W(t) via

dWT(t)=dW(t)+σP(t,T)dt,(4.1) \text{d}W^T(t) = \text{d}W(t) + \sigma_P(t,T)\text{d}t, \tag{4.1}

so that by Girsanov's theorem, the density process η(t)=EtQ[dQTdQ]\eta(t) = \mathbb{E}_t^\mathbb{Q}\left[\frac{\text{d}\mathbb{Q}^T}{\text{d}\mathbb{Q}} \right] for the measure shift between QT\mathbb{Q}^T and Q\mathbb{Q} satisfies

dη(t)η(t)=σP(t,T)dW(t). \frac{\text{d}\eta(t)}{\eta(t)} = -\sigma_P(t,T)^\intercal \text{d}W(t).

Instantaneous Forward Rate Dynamics

While the above fully specifies the HJM dynamics of the ZCB and forward bond prices, the HJM framework is normally expressed in terms of the instantaneous forward rates. Doing so eliminates the need to consider the short rate rr and also reveals a number of the fundamental properties of HJM models. By Ito's lemma, under the risk-neutral measure, the quantity Y(t,T):=logP(t,T)Y(t,T) := \log P(t,T) satisfies

dY(t,T)=σP(t,T)dW(t)+g(t,T)dt, \text{d} Y(t,T) = - \sigma_P(t,T)^\intercal \text{d}W(t) + g(t,T)\text{d}t,

for some function gg that is irrelevant for our purposes. Differentiating both sides of the above SDE by TT and taking the partial differential inside the differential term, we see that f(t,T)=Y(t,T)Tf(t,T) = \tfrac{\partial Y(t,T)}{\partial T} satisfies SDE

df(t,T)=μf(t,T)dt+σf(t,T)dW(t), \text{d}f(t,T) = \mu_f(t,T)\text{d}t + \sigma_f(t,T)^\intercal\text{d}W(t),

where

σf(t,T)=TσP(t,T).(4.2) \sigma_f(t,T) = \frac{\partial}{\partial T} \sigma_P(t,T). \tag{4.2}

Using the relationship (4.1), we find that

df(t,T)=μf(t,T)dt+σf(t,T)(dWT(t)σP(t,T)dt)=(μf(t,T)σf(t,T)σP(t,T))dt+σf(t,T)dWT(t). \begin{aligned} \text{d}f(t,T) & = \mu_f(t,T)\text{d}t + \sigma_f(t,T)^\intercal(\text{d}W^T(t) - \sigma_P(t,T)\text{d}t) \\ & = (\mu_f(t,T) - \sigma_f(t,T)^\intercal \sigma_P(t,T))\text{d}t + \sigma_f(t,T)^\intercal \text{d}W^T(t). \end{aligned}

Since f(t,T)f(t,T) is a QT\mathbb{Q}^T martingale, the drift term must vanish. Furthermore, integrating (4.2) with respect to T gives σP(t,T)=tTσf(t,u)du\sigma_P(t,T) = \int_t^T \sigma_f(t,u)\text{d}u. These two results together show that

μf(t,T)=σf(t,T)tTσf(t,u)du. \mu_f(t,T) = \sigma_f(t,T)^\intercal \int_t^T \sigma_f(t,u)\text{d}u.

This is the main result of Heath, Jarrow and Morton. Once the forward rate diffusion coefficients σf(t,T)\sigma_f(t,T) have been specified, the HJM model is fully specified. The HJM models take initial forward rates f(0,T)f(0,T) as inputs, so that the models are automatically consistent with discount bond prices at time 00. This is true irrespective of the choice of σf(t,T)\sigma_f(t,T), which can be set freely.

This automatic calibration to bond prices is the main selling point of HJM models. However, there are a number of less desirable features of HJM models. Particularly problematic is the dimensionality of the model. To describe bond prices at some future time tt, an entire continuum of forward rates {f(t,u):u[t,T]}\{f(t,u): u \in [t,T]\}. In practice, to make this tractable requires either making assumptions on σf\sigma_f that permits a finite-dimensional Markovian representation of the forward rate curve, or moving to a finite state space model that tracks forwards rates spanning time-buckets of finite length. This second approach is explored in much more depth in the Libor market model (LMM).

It's important to note that any arbitrage-free IR model set in a finite dimensional Brownian motion world must be a special case of an HJM model. In particular, any such model corresponds to a particular choice of σf(t,T)\sigma_f(t,T). So, in some sense, the HJM framework underlies all of the models that we will cover.

Short Rate Processes

In principle specification of the short-rate process is sufficient to completely specify a full yield curve model. Under the HJM framework, it follows that the short-rate r(t)r(t) under Q\mathbb{Q} satisfies

r(t)=f(t,t)=f(0,t)+0tσf(u,t)utσf(u,s)dsdu+0tσf(u,t)dW(u). \begin{aligned} r(t) & = f(t,t) \\ & = f(0,t) + \int_0^t \sigma_f(u, t)^\intercal \int_u^t \sigma_f(u,s)\text{d}s \text{d}u + \int_0^t \sigma_f(u, t)^\intercal \text{d}W(u). \end{aligned}

Clearly the process r(t)r(t) is not Markovian in general, as can be seen by focusing on the term

D(t)=0tσf(u,t)dW(u), D(t) = \int_0^t \sigma_f(u,t)^\intercal \text{d}W(u),

for which we have

D(T)=D(t)+tTσf(u,T)dW(u)+[0tσf(u,T)dW(u)0tσf(u,t)dW(u)]. \begin{aligned} D(T) & = D(t) + \int_t^T \sigma_f(u,T)^\intercal \text{d}W(u) + \left[\int_0^t \sigma_f(u,T)^\intercal \text{d}W(u) - \int_0^t \sigma_f(u,t)^\intercal \text{d}W(u) \right]. \end{aligned}

Unless the square-bracketed terms are either non-random, or a deterministic function of D(t)D(t), we find that

EQ[D(T)D(t)]EtQ[D(T)]. \mathbb{E}^\mathbb{Q}\left[D(T)\vert D(t)\right] \neq \mathbb{E}_t^\mathbb{Q}[D(T)].

Examples of HJM Models

The Gaussian Model

In the Gaussian model, we assume that σP(t,T)\sigma_P(t,T) is abounded deterministic function of tt and TT. It follows that forward bond prices are then log-normal under both Q\mathbb{Q} and QT\mathbb{Q}^T. The forward rate process under Q\mathbb{Q} is

df(t,T)=σf(t,T)σP(t,T)dt+σf(t,T)dW(t),σf(t,T)=TσP(t,T). \begin{aligned} \text{d}f(t,T) & = \sigma_f(t,T)^\intercal \sigma_P(t,T)\text{d}t + \sigma_f(t,T)^\intercal\text{d}W(t), \\ \sigma_f(t,T) & = \frac{\partial}{\partial T} \sigma_P(t,T). \end{aligned}

This implies that r(T)r(T) is Gaussian with Q\mathbb{Q} moments

EtQ[f(T,T)]=tTσf(u,T)σP(u,T)du,VartQ(f(T,T))=tTσf(u,T)σf(u,T)du. \begin{aligned} \mathbb{E}_t^\mathbb{Q}\left[f(T,T) \right] & = \int_t^T \sigma_f(u,T)^\intercal \sigma_P(u,T) \text{d}u, \\ \text{Var}_t^\mathbb{Q}(f(T,T)) & = \int_t^T \sigma_f(u,T)^\intercal \sigma_f(u,T)\text{d}u. \end{aligned}

The Gaussian short-rate model benefits from being tractable and permitting analytical price formulas for a number of European options and futures. It is often a good starting point for the development of closed-form approximations in other models. An example of a derivative admitting a closed-form solution under the Gaussian model is an option on a ZCB:

Example (Option on a ZCB)

Consider a European call option with maturity SS and strike price KK on a TT-bond. The payoff at time SS is given by

V(S)=(P(S,T)K)+,S<T. V(S) = (P(S,T) - K)^+, \quad S < T.

Under the SS-forward measure QS\mathbb{Q}^S we know that

V(t)=P(t,S)EtS[(P(S,S,T)K)+]. V(t) = P(t,S)\mathbb{E}_t^S\left[\left(P(S, S, T) - K\right)^+ \right].

Applying the dynamics of P(t,S,T)P(t, S, T) we find that

V(t)=P(t,T)ϕ(d+)KP(t,S)ϕ(d), V(t) = P(t, T)\phi(d_+) - K P(t, S)\phi(d_-),

where

d±=log(P(t,T)KP(t,S))±12σ2σσ2=tTσP(u,T)σP(u,S)2du. \begin{aligned} d_{\pm} & = \frac{\log\left(\frac{P(t,T)}{K P(t,S)}\right) \pm \frac{1}{2}\sigma^2}{\sigma} \\ \sigma^2 & = \int_t^T \left\vert \sigma_P(u,T) - \sigma_P(u,S) \right\vert^2 \text{d}u. \end{aligned}

Gaussian Models with Markovian Short Rate

Although tractable, the Gaussian model does not generally allow for a finite-dimensional Markovian representation, and does not typically imply Markov behaviour of the short rate. By considering the special choice of forward rate volatility

σf(t,T)=g(t)h(T), \sigma_f(t,T) = g(t)h(T),

where hh is a positive-valued real function and g(t)Rdg(t) \in \mathbb{R}^d, we can ensure that r(t)r(t) is Markovian. In this case, we find that

σP(t,T)=tTσf(t,u)du=g(t)tTh(u)du(4.3)\begin{aligned} \sigma_P(t,T) & = \int_t^T \sigma_f(t,u)\text{d}u = g(t) \int_t^T h(u)\text{d}u \tag{4.3} \end{aligned}

and

r(t)=f(0,t)+h(t)0tmf(t,u)du+h(t)0tg(u)dW(u).(4.4)\begin{aligned} r(t) & = f(0,t) + h(t)\int_0^t m_f(t,u)\text{d}u + h(t) \int_0^t g(u)^\intercal \text{d}W(u). \tag{4.4} \end{aligned}

We find that the term

D(t)=0tσf(u,t)dW(u)=h(t)0tg(u)dW(u)\begin{aligned} D(t) & = \int_0^t \sigma_f(u,t)^\intercal\text{d}W(u) = h(t)\int_0^t g(u)^\intercal \text{d}W(u) \end{aligned}

is Markov, since

D(T)=h(T)0Tg(u)dW(u)=h(T)h(t)D(t)+h(T)tTg(u)dW(u).\begin{aligned} D(T) & = h(T)\int_0^T g(u)^\intercal \text{d}W(u) \\ & = \frac{h(T)}{h(t)}D(t) + h(T)\int_t^T g(u)^\intercal\text{d}W(u). \end{aligned}

It quickly follows that r(t)r(t) itself is Markovian by differentiating (4.4). In fact, we have the following result.

Proposition (Markovian short rate)

In the dd-dimensional Gaussian HJM model, when (4.3)(4.3) holds, the short rate satisfies an SDE of the form

dr(t)=(a(t)ξ(t)r(t))dt+σr(t)dW(t), \text{d}r(t) = (a(t) - \xi(t)r(t))\text{d}t + \sigma_r(t)^\intercal\text{d}W(t),

where ξ(t)R\xi(t) \in \mathbb{R} and σr(t)Rd\sigma_r(t) \in \mathbb{R}^d are deterministic functions of time, and

a(t)=f(0,t)t+ξ(t)f(0,t)+0tσf(u,t)σf(u,t)du.\begin{aligned} a(t) & = \frac{\partial f(0,t)}{\partial t} + \xi(t)f(0,t) + \int_0^t \sigma_f(u,t)^\intercal \sigma_f(u,t)\text{d}u. \end{aligned}

Log-Normal HJM Models

To avoid the issue of negative forward rates that arises in Gaussian HJM models, it is tempting to consider forward rates of the form

σf(t,T)=f(t,T)σ(t,T), \sigma_f(t,T) = f(t,T)\sigma(t,T),

where σ(t,T)\sigma(t,T) is deterministic and bounded. In the TT-forward measure,

df(t,T)=f(t,T)σ(t,T)dWT(t), \text{d}f(t,T) = f(t,T)\sigma(t,T)^\intercal \text{d}W^T(t),

so that f(t,T)f(t,T) is log-normal. This leads to severe issues. Under Q\mathbb{Q} we find that forward rates explode to infinity with non-zero probability. Hence, all valuation formulas under Q\mathbb{Q} lead to ZCB prices being zero, implying obvious arbitrage opportunities.

© 2022 Will Douglas, All Rights Reserved.