Term Structure Models

1. Zero-coupon bond prices and yields

2. Vasicek model
3. Cox-Ingersoll-Ross model
4. Multifactor Cox-Ingersoll-Ross models
5. Affine models
6. Completely affine models
7. Bond risk premia
8. Inflation and nominal asset prices

Readings and References

Back, chapter 17.
Duffie, chapter 7.
Vasicek, O., 1977, An equilibrium characterization of the term structure, Journal of
Financial Economics 5, 177-188.
Cox, J., J. Ingersoll, and S. Ross, 1985, A theory of the term structure of interest
rates, Econometrica 53, 385-407.
Longsta↵, F., and E. Schwartz, 1992, Interest rate volatility and the term structure:
A two-factor general equilibrium model, Journal of Finance 47, 1259-1282.
Duffie, D., and R. Kan, 1996, A yield-factor model of interest rates, Mathematical
Finance 6, 379-406.
Du↵ee, G., 2002, Term premia and interest rate forecasts in affine models, Journal
of Finance 57, 405-443.
Drechsler, I., A. Savov, and P. Schnabl, A Model of Monetary Policy and Risk
Premia, Journal of Finance forthcoming.

2
 Summary of the Continuous-Time Financial Market
dS0,t dSk,t
I Security prices satisfy S0,t
= rt dt and Sk,t
= (µk,t k,t ) dt + k,t dBt .

I Given tight tr. strat. ⇡t and consumption ct, portfolio value Xt satisfies the
• WEE: dXt = rtXt dt + ⇡t(µt rt1) dt + ⇡t t dBt ct dt.

I No arbitrage ) if ⇡t t = 0 then ⇡t(µt rt1) = 0 ) 9✓t s.t. t ✓t = µt rt 1

) dXt = rtXt dt + ⇡t t(✓t dt + dBt) ct dt.
Rt R
1 t |✓ |2 ds
dP ⇤ ✓ 0 dBs
I Under emm P ⇤ given by dP
= ZT where Zt = e 0 s 2 0 s ,
Rt
• Bt⇤ = Bt + 0 ✓s ds is Brownian motion.
Rt
r ds
Let t =e 0 s and sdf process Mt = t Zt . Then the WEE can also be written:
⇤
• WEE*: d tXt + t ct dt = t ⇡t t dBt

• WEE-M: dMtXt + Mtct dt = Mt[⇡t t ✓tXt] dBt

RT RT
I So Xt = E⇤t { t
u
c du +
u
T
X T } = Et { t
Mu
c du
Mt u
+M
M
T
XT } if ⇡ is mtgale-gen.
t t t

I If is nonsingular, every c.plan (c, XT ) can be generated by a mtgale-gen. tr.strat.

Zero-Coupon Bond Prices and Yields

I Now let’s go through some classic models of the prices of default-free bonds.
I Let PtT be the time t price of the zero maturing at time T and ytT be its yield.
I Suppose we’re in a complete market. We can represent zero prices in various ways:
RT MT
ytT (T t) ru du
PtT ⌘ e = E⇤t {e t } = Et { }. (1)
Mt

I For the purpose of modeling bond prices and yields, we can work entirely under
the risk-neutral measure P ⇤.
I In order to characterize expected bond returns, we need to work under the true
measure P .

4
 The Vasicek Model

I In the Vasicek (1977) model, the instantaneous riskless rate rt is modeled as a

constant-volatility mean-reverting Ornstein-Uhlenbeck process,

drt = (r̄ rt) dt + dB ⇤ , (2)

where r̄ is the “long-run mean” of rt and  > 0 is its “speed of mean-reversion.”

I The Ornstein-Uhlenbeck process is the continuous-time analogue of the discrete-
time AR(1) process.
I The solution of (2) for any u t 0 is
Z u
ru = e (u t)
rt + (1 e (u t)
)r̄ + e (u s)
dBs⇤, (3)
t

so ru is normally distributed with conditional mean and variance

E⇤t {ru} = e (u t)

rt + (1 e (u t)
)r̄ and (4)
Z u 2
2 2(u s) 2(u t)
var⇤t {ru} = e ds = (1 e ). (5)
t 2
I Normality of ru means negative interest rates are possible in this model.

RT
T
I It follows from PtT ⌘ e yt (T t)
= E⇤t {e t
ru du
} that zero yields are “affine” in
rt, and thus also normal:

ytT = a(T t) + b(T t)rt, where (6)

2 2 2 ⌧ 2 2⌧
r̄ 1 e 1 e
a(⌧ ) = r̄ +( )( ) ( ) and (7)
22 3 ⌧ 4 3 ⌧
⌧
1 1 e
b(⌧ ) = ( ). (8)
 ⌧

I The derivation of this result from Back (2010) is as follows:

Z T Z T Z T Z u
ru du = (T t)r̄ (r̄ rt ) e (u t)
du + e (u s)
dBs⇤ du
t t t t
Z Z
1 T u
= (T t)r̄ (r̄ rt)(1 e (T t)
)+ e (u s)
dBs⇤ du .
 t t

I Switching the order of integration in the last integral gives

Z T Z u Z T Z T Z T
e (u s)
dBs⇤ du = e (u s)
du dBs⇤ = (1 e (T s)
) dBs⇤ .
t t t s  t

6
I The integrand of this stochastic integral is nonrandom, so the integral is normal
with mean zero and variance
2 Z 2
T
(T s) 2 2 2 (T t)
2
2(T t)
(1 e ) ds = (T t) (1 e )+ (1 e ).
2 t 2 3 23
RT
I Thus t ru du is normal with
Z T 1
E⇤t { ru du} = (T t)r̄ + (r̄ rt)(1 e (T t)) and (9)
t 
Z 2
T 2 2 2
2(T t)
var⇤t { ru du} = (T t) 2 (1 e (T t)
) + (1 e ),
t  3 23
(10)

I and it follows from the usual rule for expectations of exponentials of normals that
RT 1
ru du
log PtT = log E⇤t {e t }= (T t)r̄ + (r̄ rt)(1 e (T t)
)

2
1 2 2 (T t)
2
2(T t)
+ [(T t) (1 e )) + (1 e ))] . (11)
2 2 3 23

I Homework: Verify that PtT satisfies the fundamental PDE for derivative prices.

The Cox-Ingersoll-Ross Model

I In the Cox-Ingersoll-Ross (1985) model, the instantaneous riskless rate rt is

modeled as a mean-reverting square-root process,
p
drt = (r̄ rt) dt + rt dB ⇤ , (12)

where r̄ > 0 is the long-run mean of rt,  > 0 is its speed of mean-reversion, and
the solution to (12) is nonnegative for all t, so it is possible to take the square
2
root. If r̄ /2, then rt is a.s. strictly positive for all t.
I The exact solution to (12) is more complicated than the solution to the O-U
process in Vasicek, so we derive zero prices by solving their fundamental PDE.
I Since the square-root process (12) is Markov, zero prices are functions of rt:
RT
ru du
PtT = E⇤t {e t } = pT (rt, t) . (13)

I Assuming the price function pT is smooth enough for an application of Itô’s lemma,

1 2 @P T
(r̄ r)pTr + rpTrr + = rpT s.t. pT (r, T ) = 1 . (14)
2 @t

8
I The solution to (14) has yields affine in rt:

(T t)rt
pT (rt, t) = e ↵(T t)
and ytT = a(T t) + b(T t)rt . (15)

where a(⌧ ) = ↵(⌧ )/⌧ and b(⌧ ) = (⌧ )/⌧ and

2r̄ ( + )⌧ 2 2(e ⌧ 1)
↵(⌧ ) = 2
[ + log ] and (⌧ ) = , (16)
2 c(⌧ ) c(⌧ )
p
where c(⌧ ) = 2 + ( + )(e ⌧ 1) and = 2 + 2 2 .
I To verify this solution, note that (15) implies pr = p, prr = 2
p, and
0 0
pt = (↵ + r )p, so the fundamental PDE holds if and only if

1 2 2
(r̄ r) + r + ↵0 + 0
r=r. (17)
2
This holds for all values of r if and only if

1 2 2 0
 + + = 1 and ↵0 = r̄ . (18)
2
The ODE for above is said to be of the Riccati type.

I The terminal condition pT (r, T ) = 1 is equivalent to ↵(0) = (0) = 0.

I Di↵erentiating in (16) shows it is a solution to (18).
I Given , integrating ↵0 in (18) gives ↵ in (16).

10
 The Multifactor CIR Model

We can get a multifactor version of the CIR model by taking r to be the sum of
square-root processes: rt = X1t + X2t where
p
dXit = i(X̄i Xit) dt + i Xit dBi⇤ . (19)

I Independence of the Xi implies

RT RT RT
ru du X1u du X2u du
pTt = E⇤t {e t } = E⇤t {e t }E⇤t {e t } (20)
↵1 (T t) ↵2 (T t) 1 (T t)X1t 2 (T t)X2t
=e (21)

where the ↵i and i are defined as in the single-factor CIR model.

↵i (⌧ ) i (⌧ )
I This implies ytT = y1t
T T
+ y2t where yitT = ⌧
+ ⌧
Xit.

Other Multifactor Models Similarly, we could construct other multifactor interest

rate processes as sums of independent processes, and get product pricing formulas if
price formulas exist for each independent process.

11

Affine Term Structure Models

In an n-factor affine term structure model, the short rate r is affine in n Markov state
variables X whose drift and instantaneous covariance matrix are also affine in X :

rt = 0 + 0Xt and dXt = ( + KXt) dt + (Xt) dB ⇤, where (22)

I B ⇤ is a vector of n independent Brownian motions under P ⇤,

I 0 is a constant, and are a constant n-vectors, K is a constant n ⇥ n matrix,
I is an n ⇥ n matrix-valued function s.t. each element of the covariance matrix
(x) (x)0 is affine in x. Parametric assumptions are made to ensure the diagonal
elements of (x) (x)0 are nonnegative and uniqueness of the solution for X .
I Both the Vasicek and CIR models are examples of affine models.
I If is constant, then the model is Gaussian, in the sense that conditional on Xt,
(ru, Xu) is multivariate normal for all u t.
I It can be shown that in any two-factor Gaussian model, the two factors can be
taken to be the short rate and its drift.

12
I In an affine model, zero prices are exponential-affine in X and yields are affine:

Pn n
X
pT (x, t) = e ↵(T t) i=1 i (T t)xi
and ytT = a(T t) + bi(T t)xi , (23)
i=1

as shown by verifying the fundamental PDE for p with BC ↵(0) = i (0) = 0.

I The i solve a multi-dimensional ODE, for which no closed-form is available in

general, and ↵ can be solved by integrating the i.
I A useful feature of affine models is that one can take a vector of zero yields to be
the factors. For the vector Y of n yields for fixed times to maturity ⌧1, . . . , ⌧n, let
A be the vector of the a(⌧i) and B the matrix of the bj (⌧i). Then

1
Yt = A + BXt ) Xt = B (Yt A) , (24)

provided B is invertible. Substituting Xt = B 1(Yt A) and dX = B 1

dYt
produces an affine model with the n yields as factors.
I Similarly, one could use n 1 zero yields and the short rate as affine factors.

13

I To get the fundamental PDE, write PtT = e Yt

where Y = ↵(T t)+ (T t)>Xt
I From Itô’s lemma
dP 1
= dY + (dY )2 = ↵0(T t) dt + 0(T t)>X dt (T t)> dX
P 2
1
+ (T t)>(dX)(dX)> (T t) dt (25)
2
= ↵0(T t) dt + 0(T t)>X dt (T t)>[( + KX) dt + (X) dB ⇤]
1
+ (T t)> (X) (X)> (T t) dt . (26)
2
I Setting the appreciation rate of P equal to the short rate r = 0 + 0X gives:

0 + 0X = ↵0(T t) + 0
(T t)>X (T t)>( + KX) (27)
1
+ (T t)> (X) (X)> (T t) . (28)
2
I Equating the coefficients of the Xi on each side gives a system of n ODEs in
the functions i, called Riccati equations because they are affine in the i0 and
quadratic in the i.
I Matching constant terms on each side, given the i, determines ↵0, which can be
integrated from ↵(0) = 0 to give ↵.

14
 Completely Affine Models

In a completely affine model, factor dynamics are affine under both the risk-neutral
measure P ⇤ and the true measure P . This is set up as follows.
I Assume (X) in the affine model (23) is of the form S(X) where is a
constant n ⇥ n matrix and S is a diagonal n ⇥ n matrix-valued function with each
squared diagonal element being an affine function of X . Then each element of
the covariance matrix (x) (x)0 = S(x)2 0, is an affine function of x.
dP ⇤
I Assume the market price of risk ✓ in the Radon-Nikodym derivative dP
= Z(T ) =
RT R
1 t
✓ dBu 2 0 |✓u |2 du
e 0 u is of the form ✓t = S(Xt)✓ for a constant vector ✓.
I Then dBt⇤ = dBt + S(X)✓ dt and the model is also affine under P :

dXt = ( + KXt) dt + S(Xt) dB ⇤ (29)

= [ + KXt + S(Xt)2✓] dt + S(Xt) dBt (30)
= ( ˆ + K̂X) dt + S(Xt) dBt (31)

for some constant vector ˆ and matrix K̂ , because S(X)2 is diagonal with affine
functions of X on the diagonal.

15

Bond Risk Premia

In a completely affine model, bond risk premia are also affine in X .

I To see this, note from (26) that the volatility vector of the zero price P T is
(T t)0 S(Xt).

I Thus, the risk premium on the (T t)-year zero is

µTt rt = (T t)0 S(Xt)2✓ , (32)

which is affine in Xt.

I In a Gaussian completely affine model, S(X) is a constant matrix, so the risk
premium of a zero depends only on its time to maturity.

16
 Inflation and Nominal Asset Prices

I Until now, we have been working with real prices Sk,t, in units of the consumption
good, and the sdf Mt for these real prices. When we go to the data, we can in
some contexts ignore inflation, but it’s an issue when pricing nominal bonds, i.e.,
claims to a dollar rather than a unit of consumption.
I Let qt be the price level, e.g., dollars per unit of consumption, sometimes measured
by CPI.
I Then nominal asset prices are qtSk,t.
I Therefore, the sdf for nominal asset prices is Mtq ⌘ Mt
qt
, since MtSt = Et{MuSu}
implies Mtq qtSt = Et{Muq quSu}.
Rt Rt R
1 t| 2
I Suppose dqt
qt
= ◆t + q,t dBt , or qt = e ◆ ds+ 0 q,s dBs
0 s 2 0 q,s | ds .

I Interpret ◆t as expected inflation over the period and dqt

qt
as realized inflation.

17

Locally Riskless Inflation

I If q,t = 0, i.e., inflation is locally riskless or “known at the beginning of the

period,” then nominal asset prices qtSk,t follow

dqtSk,t
= (µk,t + ◆t k,t ) dt + k,t dBt , (33)
qtSk,t
Rt Rt 0 1
Rt 2
I the sdf for nominal prices is Mtq = Mq t = e 0 (rs+◆s) ds 0 ✓s dBs 2 0 |✓s| ds, and
t
I the nominal riskless rate, i.e., the nominal return on a locally riskless nominal
money market account is rt + ◆t.
I In this case, nominal excess expected returns (µk,t + ◆t) (rt + ◆t) are equal to
real excess returns µk,t rt, so theory developed for real excess returns may be
able to be applied directly to empirical work with nominal excess returns.
I The price of a “nominal bond” paying a dollar at time T is

MTq RT
(rs +◆s } ds
Ptq,T = Et{ 1} = E ⇤
t {e t }. (34)
Mtq

If both rt and ◆t are affine in a set of affine state variables,

Pn then bond prices will
q,T ↵(T t) (T t)xi
again be exponentially affine, p (x, t) = e i=1 i .

18
 Inflation with Shocks

I If q,t 6= 0, i.e., if realized inflation over the period contains both its expectation
◆t dt and a shock q,t dBt, then nominal asset prices qtSk,t follow

dqtSk,t dSk,t dqt dq dSk

= + + ( )( ) (35)
qtSk,t Sk,t qt q Sk
0
= (µk,t + ◆t + k,t q,t k,t ) dt +( k,t + q,t ) dBt , (36)

I the sdf for nominal prices is

Mt Rt Rt 0 R R
1 t |✓ |2 ds+ 1 t | 2
Mtq = = e 0 (rs+◆s) ds 0 (✓s+ q,s ) dBs 2 0 s 2 0 q,s | ds , (37)
qt
Rt Rt R
t
(rs 2
q,s ✓s +◆s | q,s | ) ds (✓s0 + q,s ) dBs 1 0 2
2 0 |✓s + q,s | ds
=e 0 0 , (38)

I and it follows from the form of Mtq that the nominal riskless rate, the rate on a
locally riskless nominal money market account, is rt q,t ✓t + ◆t | q,t|2.

19

The Nominal Riskless Rate and the Inflation Risk Premium

I To get to this more intuitively, note that the nominal money market account,
which delivers a locally riskless nominal return and zero shocks, must have shocks
to its real returns that are exactly opposite the shocks to the price level. In other
words, if the nominal money market is security k⇤, with real price Sk⇤ , its real
return shocks must by equation (36) have loadings k⇤ identically equal to q.

I Thus the nominal money market’s real risk premium µk⇤,t rt must be k⇤ ✓ =
q ✓ , which it must pay because its real return in units of consumption is risky.

I As we will see shortly, this q ✓ gives rise to the so-called “inflation risk premium”
in the term structure of nominal bond yields.
I As we will see later, in equilibrium q ✓ is the instantaneous covariance between
shocks to inflation and shocks to aggregate consumption, times the relative risk
aversion of the representative agent. Estimates suggest that this has been positive
historically, though perhaps not since the crisis.

20
I Now to get the nominal riskless rate, note that by equation (36) with k⇤ = q,
the nominal money market account’s nominal price follows

dqtSk⇤,t 0
= (µk⇤,t + ◆t + k⇤ ,t q,t ) dt +( k⇤ ,t + q,t ) dBt (39)
qtSk⇤,t
2
= (rt q,t ✓t + ◆t | q,t | ) dt . (40)

I Therefore, the nominal riskless rate is rt q,t ✓t + ◆t | q,t |

2
.
I To gain economic intuition for all this, let’s redefine inflation as the rate of decline
in the consumption price of dollars 1q , rather than the rate of increase in the dollar
price of consumption q or CPI, that is more typically quoted in practice. Viewing
currency (e.g., dollars) as an asset priced in units of consumption will clarify the
economics of its real return and the real returns of nominal assets more generally.
I By Itô’s lemma,

1 1 12 1 1 2
d = 2
dqt + (dq) = (◆t dt + q,t dBt ) + | q,t | dt (41)
qt qt 2 qt3 qt qt

21

I so d 1/q
1/qt
t
= ( ◆t +| q,t|2) dt q,t dBt = it dt+ 1/q,t dBt, where it = ◆t | q,t|2
is the rate of decline in the real price of dollars, our alternative measure of inflation,
and 1/q,t = q,t is the volatility of the real price of dollars.

I Then we can see the nominal riskless rate as rt + it + 1/q,t✓t, the real riskless
rate on rate rt plus compensation for the decline in the real price of dollars it plus
compensation for the risk of the real price of dollars 1/q,t✓t.
I If shocks to inflation and aggregate consumption are positively correlated, i.e.,
q ✓ > 0, then the nominal money market is a hedge against negative consumpion
shocks, because it pays dividends in dollars, which are worth more in real terms,
when inflation and consumption are down. In that case, it makes sense that the
real risk premium on the nominal money market is negative, 1/q ✓ = q ✓ < 0.

22
 Nominal Bond Prices

I The nominal price of the nominal zero paying a dollar at time T is

MTq RT RT T R
2 (✓s0 + q,s ) dBs 1 0 2
Ptq,T = Et{ q 1} = Et{e t (rs q,s ✓s +◆s | q,s | ) ds t 2 t |✓s + q,s | ds }
Mt
(42)
RT 2
(rs q,s ✓s +◆s | q,s | ) ds
= Eqt {e t } (43)

under the risk-neutral measure for nominal prices P q given by

dP q Rt 0 R
1 t |✓ + 0 |2 ds
= e 0 (✓s+ q,s ) dBs 2 0 s q,s , (44)
dP
Rt
under which Btq ⌘ Bt + 0
0 (✓s + q,s ) ds is a standard Brownian motion.
I Therefore, if rt and ◆t are affine in a set state variables Xt that are affine under
P q , and if q,t = q S(Xt) and ✓t = S(Xt)✓ , where q and ✓ constant vectors
and S(X) a diagonal matrix as in the completely affine model, then the nominal
riskless rate rt q,t ✓t + ◆t | q,t|2 will be affine in the state variablesPand nominal
n
zero prices will again be exponentially affine, pq,T (x, t) = e ↵(T t) i=1 i(T t)xi .

23

Nominal Risk Premia

I Note that under P q , nominal expected returns on all assets are equal to the
nominal riskless rate, as can be seen by substituting dBtq (✓t + q,t) dt for dBt
in equation (36):

dqtSk,t 0
= (µk,t + ◆t + k,t q,t k,t ) dt +( k,t + q,t ) dBt (45)
qtSk,t
2 q
= (rt q,t ✓t + ◆t | q,t | k,t ) dt +( k,t + q,t ) dBt . (46)

This justifies calling P q a risk-neutral measure for nominal returns.

I Nominal risk premia, expected nominal returns minus the nominal riskless rate, are

0 2
(µk,t + ◆t + k,t q,t ) (rt q,t ✓t + ◆t | q,t | ) (47)
2 0
= µk,t rt + q,t ✓t +| q,t | + k,t q,t (48)
q 0
= k,t (✓t + q,t ) (49)

q
where k,t ⌘ k,t + q,t is the vector of security k’s nominal return risk loadings,
0
and ✓ + q is the mpr paid by nominal returns.
I Now the nominal risk premia di↵er from the real risk premia µk,t rt.

24