Asia-Pacific Financial Markets 10: 59–85, 2003.

C 2004 Kluwer Academic Publishers. Printed in the Netherlands.

Prediction of Individual Bond Prices via a Dynamic

Bond Pricing Model: Application to Japanese
Government Bond Price Data

HIROSHI TSUDA
Financial Research Group, The NLI Research Institute, 1-1-1 Yuraku-Cho Chiyoda-ku,
Tokyo 100-0006, Japan, e-mail: tsuda@nli-research.co.jp

Abstract. In this paper, we propose a dynamic bond pricing model and report the usefulness of our
bond pricing model based on analysis of Japanese Government bond price data. We extend the concept
of the time dependent Markov (TDM) model proposed by Kariya and Tsuda (Financial Engineer-
ing and the Japanese Markets, Kluwer Academic Publishers, Dordrecht, The Netherlands, Vol. 1,
pp. 1–20) to a dynamic model, which can obtain information for future bond prices. A main feature of
the extended model is that the whole stochastic process of the random cash-flow discount functions
of each individual bond has a time series structure. We express the dynamic structure for the models
by using a Bayesian state space representation. The state space approach integrates cross-sectional
and time series aspects of individual bond prices. From the empirical results, we find useful evidence
that our model performs well for the prediction of the patterns of the term structure of the individual
bond returns.

Key words: Bayesian state space representation, dynamic bond pricing model, Kalman filter, random
cash-flow discount function

1. Introduction
We propose a dynamic bond pricing model to express the time series and cross-
sectional structure of individual bond prices. The framework of our model includes
the time dependent Markov (TDM) model proposed by Kariya and Tsuda (1994),
which is a cross-sectional model of individual bond prices, but which does not
include a time series structure to predict individual future bond prices. Our model
is different from the TDM model in that we employ both of a time series and
cross-sectional structure of individual bond price movements. On the other hand,
our model is also quite different from many other normative bond pricing models
modified for interest rate derivatives in that it directly models the random feature
of coupon bond prices by the process of a random discount function (not by the
formulation of spot and forward rate processes), and also prices all individual bonds
having different attributes.
In addition, we propose a Kalman filtering approach based on a Bayesian state
space model to estimate our dynamic bond pricing model and to predict indi-
vidual bond prices. While the Kalman filtering approach is already traditional in
60 HIROSHI TSUDA

econometrics, it is still relatively new but growing in finance. In the literature, Babbs
and Nowman (1999), and Geyer and Pichler (1999) employ the Kalman filtering
approach for the estimation of the generalized Vasicek model and the K-factor’s
Cox-Ingersoll-Ross model on interest rates, respectively.
We empirically tested our model with Japanese Government bond (JGB) data to
investigate its usefulness. Our model performs very well as it stands. Our model is
firstly estimated with the data sample period from January 1980 to December 1989
of monthly individual bond prices and predicts their prices at the end of January
1990, after which the data sample period is extended monthly until November
2002. In fact, the standard deviation of predicted price errors for the period 1990.1
through 2002.12 is 0.96 yen on the average, where the face value of a JGB is around
100 yen. Among 156 (12 month × 13 years) months tested from 1990.1 through
2002.12, there were only 6 months with standard deviations of predicted price errors
greater than 3.0 yen. In particular, our model well captures the patterns of the term
structure of the realized returns of individual bonds. Moreover, we compare the
predictive performance of our model with an approach based on the TDM model as
follows. First, we estimate the parameters of the TDM model at each time t. Next,
we calculate individual bond prices of time t + 1 by predicting the parameters for
which the model is assumed to have the same time series structure as our model.
We found that the cumulative investment profit of the portfolio consisting of bonds
chosen by our model was larger than the one chosen by the approach based on the
TDM model. Hence, the information of our model has practical benefits for the
construction of bond portfolios and bond trading.
The paper is organized as follows. In Section 2, we first summarize the concept
of the TDM models proposed by Kariya and Tsuda (1994). Then we propose a
dynamic bond pricing model to express the time series and cross-sectional structure
of individual bond prices. Next, we present the estimation method of our model
based on a Bayesian state space representation by using a Kalman filtering approach.
In Section 3, we report the empirical results of the predictive performance of the
models, where the usefulness of our model based on analysis of JGB price data is
shown. In Section 4, we present our conclusion.

2. Bond Pricing Model

2.1. THE CONCEPT OF TIME DEPENDENT MARKOV MODEL

First, we provide a brief description of the concept of TDM model proposed by

Kariya and Tsuda (1994), from which our model is derived. Individual bond prices
are generated by investor preferences for various bond characteristics. The bond
characteristics include not only directly observable prior issue attributes such as
coupon rate and term to maturity but also unknown issue attributes which are
impossible to identify in advance. The TDM model uses the past characteristics of
the discount function as the proxy variables of the unknown issue attributes.
PREDICTION OF INDIVIDUAL BOND PRICES 61

Here a bond is defined as fixed income securities that promise to pay out a cash
flow at a certain interest rate (coupon rate) and predetermined intervals. At present
time t, we assume that a coupon bond promises to pay an amount C(s j ) at M
intervals of t + s j ( j = 1, . . . , M). For a coupon bond that makes two payments a
year, with a coupon rate of c%, and maturity s M , we have:

C(s1 ) =C(s2 ) = · · · = C(s M−1 ) = 0.5c,

C(s M )=100 + 0.5c. (1)

To model the price fluctuation structure for a coupon bond i at time t, the TDM
model expresses the stochastic process of the discount function Dit (s), which is
dependent on issue attributes. To note that cash flows are generated at different
times for different issue attributes of the bonds, the term (period) from time t till
the j-th cash flow of the i-th bond is expressed as

s j = s(i) j ( j = 1, . . . , M(i) : i = 1, . . . , N ). (2)

All these terms in N bonds are enumerated in increasing order as

sa1 < sa2 < · · · < sa M , sa M = max s(i) M(i) . (3)

Hence sak is one of the terms in which at least one of the N bonds yields a cash
flow. Let Cit (s) be the cash flow function of the i-th bond defined on 0 ≤ s ≤ sa M .
Of course, Cit (s) = 0 unless s = sak = s(i) j for some j. In the model, the
realization of price Pit (0), which is random, is viewed as equivalent to a whole
realization of the stochastic process

Dit = Dit (s) : 0 ≤ s ≤ sa M (4)

of the discount function of the i-th bond in the model, where

M
Pit (0)= Cit (sa j )Dit (sa j ) = C it D it ,
j =1

C it =(Cit (sa1 ), . . . , Cit (sa M )) , D it = (Dit (sa1 ), . . . , Dit (sa M )) . (5)

Then letting D it (sa j )=E[Dit (sa j )] be the mean discount function of the i-th bond,
the model (5) becomes

Pit (0)=C it D it + ηit ,

D it =E(D it ),
ηit =C it ν it , ν it = D it − D it . (6)
62 HIROSHI TSUDA

It should be noted that C it D it is the sum of cash flows discounted in (6) by the
mean discount function, and ηit is the sum of the cash flows randomly discounted
by the random discount factor ν it .
From the viewpoint that the realization of Pit (0) is equivalent to the realization
for the stochastic term structure Dit of the discount function, the time series structure
of Pit (0) is equivalent to that of Dit , where t is treated as discrete. A random relation
between Dit and Di,t−h corresponds to a random evolution of the stochastic term
structures of random discount rates (functions) from t − h to t. Two TDM models
were proposed by Kariya and Tsuda (1994). The first model of TDM model is one
with random discount factor νit (s) = Dit (s) − D it (s) which takes into account the
stochastic effect of Dit−h on Dit via the past deviation νi,t−h . We call this the TDM
model 1. The dependence of νit (s) on νi,t−h (s) is specified by

νit (s) = ξt νi,t−h (s + h) + ωit (s), (7)

or equivalently

Dit (s) = D it (s) + ξt νi,t−h (s + h) + ωit (s), (8)

where νit (s) = Dit (s) − D it (s) as before and ωit (s) is a white noise. Here the
relation between t − h and s + h in νit−h (s + h) describes the fact that s years
from t is equal to s + h years from t − h. It is noted that ξt does not depend on
i, implying that it is a cross-sectionally invariant parameter even though it is time
dependent. To identify the starting time points of the periods till the occurrences of
cash flows, we use the new notation s at = (sa1t
, . . . , sat M ) which is a vector of cash
flow of time t.
    t        t   
D it s at = Dit sa1 , . . . , Dit sat M , D it s at = D it sa1 , . . . , D it sat M
    t        t   
ν it s at = νit sa1 , . . . , νit sat M , ω it s at = ωit sa1 , . . . , ωit sat M . (9)

The discount function of (8) is expressed as

       
D it s at = D it s at + ξt ν i,t−h s at + h1 + ω it s at , (10)

where sat j + h1 = sat−h

j and

   
ν i,t−h s at + h1 = ν i,t−h s at−h . (11)

Therefore, noting Ci,t−h (s + h) = Cit (s) and using the notation

    t   
C it s at = Cit sa1 , . . . , Cit sat M ,
PREDICTION OF INDIVIDUAL BOND PRICES 63

by (5), (10) and (11), the TDM model 1 becomes

   
Pit (0)=C it s at D it s at
       
=C it s at D it s at + ξt ηi,t−h s at−h + εit s at , (12)

where
   
C i,t−h s at−h =C it s at ,
     
ηi,t−h s at−h =C i,t−h s at−h ν i,t−h s at−h , (13)
     
εit s at =C it s at ω it s at .

The important point of the TDM model 1 is that ηi,t−h (s at−h ) expresses unknown
issue attributes which are impossible to be identified in advance.
For the specifications of the mean discount function in (12), the case is considered
where the coefficients in the polynomial depend on the characteristics of bond
attributes to be incorporated into the model:


p
D it (s) =1 + δkt (z it ) s k ,
k =1

q
δkt (z it )= δklt z ilt , (14)
l =1

where z it = (z i1t , . . . , z iqt ) is a vector of the attribute variables of the i-th bond.
The structures of the variance and covariance of ηit ’s in (6) and εit ’s in (12)
are specified to express the characteristic structures of bond price movement as
mentioned below:
(1) the price structures between a short term bond and a long-term bond are less
correlated,
(2) random discount factors for cash flows occurring close by are more correlated,
and
(3) bond prices with shorter maturities fluctuate less.
Hence, the structure of the variance and covariance of εit ’s is specified as

Cov(εit , εkt )=λikt C it Φikt C kt ,

f ikt =C it Φikt C kt , F t = ( f ikt ) (15)
where,

σ 2 aiit (i = k)
λikt = , (16)
σ 2 ρaikt (i = k)
aikt =b(z it , z kt ) exp(−|si M(i) − sk M(k) |),
64 HIROSHI TSUDA

z it =(z i1t , . . . , z iqt ) ,

Φikt =(φikt· jr ) = (exp(−|sa j − sar |)). (17)
b(z it , z kt ) is a known function of z it and z kt . The structure of the variance and
covariance of ηit ’s is the same as the one of εit ’s.
The second model of the TDM model is one with the mean discount function
D it (s at ) = E[D it (s at )], which may be viewed as the conditional mean discount
function with given past stochastic discount function νi,t−h ’s. We call this the TDM
model 2. Hence, this model can be expressed as the model (6) with the mean
discount function (14), where z iqt = ηi,t−h (s at−h ) and the other attributes z ilt ’s are
non-random. In the TDM model 2, ηi,t−h (s at−h ) is regarded as an attribute of the
i-th bond at t which may not be identified in advance.

2.2. DYNAMIC BOND PRICING MODEL

In this section, we present a dynamic bond pricing model which can obtain infor-
mation on individual future bond prices. In principle, the bond pricing models in the
TDM model cannot obtain information on individual future bond prices directly be-
cause a time series structure is not included in the parameters of the models. Hence
we propose a dynamic bond pricing model of which the parameters have a time se-
ries structure. We employ a Bayesian approach with Kalman filter to analyze the re-
lationship of the individual bond prices. Our approach follows a Bayesian state space
model on cross-sectional and time series data with stochastic parameter variation.
In our dynamic bond pricing model, we consider all terms in N bonds as gen-
eration terms of cash flow. Let Cit (s) be the cash flow function of the i-th bond
defined on 0 ≤ s ≤ sa M in (3). In the model, the realization of price Pit (0), which
is random, is viewed as equivalent to a whole realization of the stochastic process
Dit of the discount function of the i-th bond in the model. Accordingly, our model
is expressed as

M
     
Pit (0) = Cit (sat j )Dit sat j = C it s at D it s at ,
j=1 (18)
    t     t   t   t 
C it s at = Cit sa1 , . . . , Cit sat M , D it s a = Dit sa1 , . . . , Dit sa M .

For the discount function Dit (s), we assume that a random discount factor ωit (sat )
is specified to express the characteristic structures of bond price movement and the
coefficients δkt in the polynomial depend on the characteristics of bond attributes
to be incorporated into the model:
 t  t
p
 k

q
Dit sa =1 + ωit sa + δkt (z it ) sat , δkt (z it ) = δklt z ilt ,
k=0 l =1
     
εit s at =C it s at ω it s at , (19)
PREDICTION OF INDIVIDUAL BOND PRICES 65

where εit (s at ) is the sum of the cash flows randomly discounted by the random dis-
count factor ω it (s at ), and z it = (z i1t , . . . , z iqt ) is a vector of the attribute variables
of the i-th bond.
We specify the same structure of variance and covariance of εit ’s as the TDM
model in (15) and (17) to express the characteristic structures of bond price move-
ment. For the specifications of the time series structure on individual bond prices,
we assume a time series structure for the parameters of the discount function as
follows:


r
δklt = φm δkl,t−m + vklt , (20)
m =1

where φm (m = 1, . . . , r ) is the coefficients in a time series structure of the param-

eters. We specify the variance and covariance of vklt ’s as mentioned below:
     2 
v01t 0 τ01 0 0 0
 v02t   0   0 τ 2 0  
     0 

 .  ∼ N  .  ,   02
 . (21)
. 0  
 ..  ..
 ..   0 0 
v pqt 0 0 0 0 τ pq2

Basically our model is a dynamic version of the TDM model. If we assume that
z i1t = ωi,t−h as one of the attribute variables of the i-th bond and δ0lt = 0(l ≥ 2),
δk1t = 0 (k ≥ 1) in (19), our model becomes a dynamic version of the TDM model
1. And if we assume that z i1t = ωi,t−h as one of the attribute variables of the i-th
bond and δ0lt = 0 (l ≥ 1) in (19), our model becomes a dynamic version of the
TDM model 2. By predicting the parameters for the discount function at t + 1 time
as follows:


r
E(δkl,t+1 ) = φm δkl, t+1−m , (22)
m =1

we can obtain information on individual future bond prices,

M
    
E(Pi, t+1 (0))= Ci, t+1 sat+1
j E Di, t+1 sat+1
j
j=1
    
=C i, t+1 s at+1 E D i, t+1 s at+1 ,

p
E(Di,t+1 (s))=1 + E(δk,t+1 (z i, t+1 ))s k , E(δk, t+1 (z i, t+1 ))
k=0

q
= E(δkl, t+1 )z il, t+1 , (23)
l =1
66 HIROSHI TSUDA

where z il,t+1 ’s are the attribute variables of the i-th bond which are determinable
in advance at t time to avoid their prediction.

2.3. ESTIMATION OF THE MODEL

We explain an estimate method of the dynamic bond pricing model (18) ∼ (21). To
express the movements of individual bond prices, we consider a time series struc-
ture for the parameters δklt of the model. We assume that z i1t = ωi, t−h as one of the
attribute variables of the i-th bond and δ0lt = 0 (l ≥ 2), δk1t = 0 (k ≥ 1) in (19).
We define the state vector expressing unobservable parameters of the model as x t =
, δ 1, t−m , . . . , δ p,t−m ) .
(δ01t , δ 1t , . . . , δ pt , δ01, t−1 , δ 1, t−1 , . . . , δ p, t−1 , . . . , δ01, t−m
If we express the individual bond variables yit = Pit (0) − M t
j = 1 C it (sa j ) at t time

as y t = (y1t , . . . , y N t ) , the bond pricing model with a time series structure can be
expressed in state-space model form as

x t =Fx t−1 + Gv t
y t =H t x t + w t . (24)

We can assume various time series structures for the unknown parameters
δ01t , δ kt (k ≥ 1). If we assume that they exhibit a smooth stochastic movement,
they can be expressed as mth-order and nth-order stochastically perturbed differ-
ence equations,

m δ01t =v0t ,
n δ kt =v kt (k ≥ 1). (25)

Here,  denotes the difference operator defined by δ kt = δ kt − δ k, t−1 . F and G

take the following forms:
   
F1 0 G1 0
F= , G= . (26)
0 F2 0 G2

On the other hand, H t are defined by

H t = [ H 1t H 2t ]. (27)

Here, coefficient vectors F 1 , G 1 and H 1t are associated with the unknown param-
eters δ01t of the discount functions, while F 2 , G 2 and H 2t are associated with δ kt .
Typically, for m = 1 and n = 2, they are given by
   
F1 = 1 , G 1 = 1 ,
 
H 1t = ε̂t−h , (28)
PREDICTION OF INDIVIDUAL BOND PRICES 67
   
2I p(q−1) −I p(q−1) I p(q−1)
F2 = , G2 = ,
I p(q−1) 0 0
 
H 2t = U t 0 , (29)

where, I p(q−1) is a p(q − 1) × p(q − 1) unit matrix,

U t =(u1t , . . . , u N t ) , uit = (ui1t , . . . , ui pt ) ,

uir t =(u i2r t , . . . , u iqr t ),

M
 r  
u ikrt = z ikt sat j Cit sat j ,
j=1
    
ε̂t−h = ε̂1,t−h s at−h , . . . , ε̂ N ,t−h s at−h . (30)

The variance and covariance of system noise v t = (v0t , v kt , ) , and observation noise
w t = (w1t , . . . , w N t ) are distributed as
    
v0t 0
v   0  
 kt     
    Q 0 
 w1t  ∼ N  0  , , (31)
    0 Rt 
 ..   ..  
 .   .  
wN t 0

where the variance covariance matrices Q and Rt are given by

 
τ12
2
0 ··· ··· ··· ··· 0
 .. 
 0 . 0 ··· ··· ··· 0 
 
 τ1q
2
··· ··· 
  0 0 0 0 
τ01
2
0  
Q = , Qk = 
 0 ··· 0
..
. 0 ··· 0  ,
0 Qk  
 0 ··· ··· 0 τ p2
2
0 0 
 
 .. 
 0 ··· ··· ··· 0 . 0 
0 ··· ··· ··· ··· 0 τ pq
2
 
a11t f 11t ρa12t f 12t ··· ρa1N t f 1N t
 ρa21t f 21t a22t f 22t ··· ρa2N t f 2N t 
 
Rt =σ 2  .. .. .. .. , (32)
 . . . . 
ρa N 1t f N 1t ρa N 2t f N 2t ··· aN N t f N N t
f ikt =C it Φikt C kt ,
68 HIROSHI TSUDA

aikt =min(si M(i) , sk M(k) ) exp(−|si M(i) − sk M(k) |),

Φikt =(φikt· jr ) = exp(−|sa j − sar |)). (33)

Given the state space and hyper-parameters θ = (τ01 2

, τ12
2
, . . . , τ pq
2
, σ 2 , ρ),
the state variable x t can be estimated by using the Kalman filter (see Appendix). The
state space model and the Kalman filter do yield a very effective method for the
computation of the likelihood of the time series models (Caines and Rissanen, 1974;
Akaike, 1978). Let Y T = (y 1 , . . . , y T ) denote a set of observations distributed
in accordance with an assumed joint parametrically expressed density function
f T (y 1 , . . . , y T |θ) where θ is parameter vector of the distribution. The likelihood
can be expressed by using the conditional distributions as follows:

L(θ)= f T (y 1 , . . . , y T |θ)
T
= gt (y t |y 1 , . . . , y t−1 , θ)
t =1

T
= gt (y t |Y t−1 , θ), (34)
t =1

where

f t (y 1 , . . . , y t |θ)= f t−1 (y 1 , . . . , y t−1 |θ)gt (y t |y 1 , . . . , y t−1 , θ),

f 1 (y 1 |θ) ≡g1 (Y 0 , θ), Y0 = φ (35)

is obtained for t = T, T − 1, . . . , 2. The individual terms in the last expression are

given by
 
1 1  −1
gt (y t |Y t−1 , θ) = exp − et γ t|t−1 et , (36)
(2π) N /2 |γ t|t−1 |1/2 2

with et = y t − H t x t|t−1 , γ t|t−1 = H t V t|t−1 H t + Rt . Therefore, the log-likelihood

of the model is given by,

l(θ)=log L(θ)
NT 1
T
1
T
=− log 2π − log |γ t|t−1 | − e γ −1 et . (37)
2 2 t =1 2 t = 1 t t|t−1

By maximizing the log-likelihood by an appropriate numerical optimization pro-

cedure, we can obtain the maximum likelihood estimate θ of θ. By minimizing the
Akaike information criterion (AIC), we can select the best model orders.
PREDICTION OF INDIVIDUAL BOND PRICES 69

Figure 1. Number of individual bonds.

3. Empirical Analysis
3.1. PREDICTION OF BOND PRICES

In this section, we estimate and predict individual bond prices with our dynamic
bond pricing model and present some significant empirical results. The Japanese
Government bonds (JGBs) in our analysis are coupon bonds with a maturity of 10
years, and a face value of 100 yen. We use the bond attributes and monthly (end-
of-month) data for bond prices from January 1980 to December 2002. At each time
point of analysis, there are 35 to 87 bonds with different coupon rates, maturities
and issue dates (Figure 1). The JGB market is an OTC (over-the-counter) market,
and our data is compiled by Nikkei and maintained at the NLI Research Institute.
Now we present the results for comparison of the models. We assumed that
the discount functions Dit (s) of the model are a second order polynomial of the
durations to cashflows. As first one of the attributes for the discount function, we
chose z i1t = ωi,t−1 (s at−1 ) at t − 1 time, which expresses an unknown issue attribute
which is impossible to be identified in advance. A second attribute z i2t is a coupon
rate and a third attribute z 13t is a term to maturity, because the model selecting those
attributes in the discount function performs very well in Kariya and Tsuda(1994).
We assume δ0lt = 0 (l ≥ 2), δk1t = 0 (k ≥ 1) to compare with the TDM model
1, because Kariya and Tsuda reported that this is the best model in terms of the
empirical results. Hence the discount function in (19) is specified as follows:

Dit (s) = 1 + ωit (s) + δ01t z i1t + (δ12t z i2t + δ13t z i3t )s + (δ22t z i2t + δ23t z i3t )s 2 .
(38)
70 HIROSHI TSUDA

We set h = 1 in (30). And we can assume various time series structures for the
parameters of the discount function. Here we assume the time structures (25) for
the parameters of the discount function are as follows,
type 1. δ01t and δ kt (k ≥ 1) are specified as a first order stochastically perturbed
difference equation,
type 2. δ01t and δ kt (k ≥ 1) are specified as a first order and a second order
stochastically perturbed difference equation, respectively.
For example, the model of (24) in type 2 form of the time structure is specified as:
   
δ01t δ01,t−1
 δ12t   δ12,t−1 
   
   
 δ13t   δ13,t−1 
      
 δ22t  1 0 0  δ22,t−1  1 0  
   
 δ23t = 0 2I 4 −I 4   δ23,t−1  +  0 I 4  v0t
    v kt
δ  δ 
 12,t−1  0 I4 0  12,t−2  0 0
   
 δ13,t−1   δ13,t−2 
   
 δ22,t−1   δ22,t−2 
δ23,t−1 δ23,t−2
 
δ01t
 δ12t 
 
 
     δ 13t 
y1t ε̂1,t−1 (s at−1 ) u 121t u 131t u 122t u 132t 0 · · · 0   δ22t 

 ..   
..   δ 
 .  =
.. .. .. .. .. .. 
. . . . . . .   23t  + w t .
 
yN t ε̂ N ,t−1 (s at−1 ) u N 21t u N 31t u N 22t u N 32t 0 · · · 0  δ12,t−1 
 
 δ13,t−1 
 
 δ22,t−1 
δ23,t−1
(39)
Here I 4 is a 4 × 4 unit matrix. The system noise v0t , v kt and observation noise w t
are distributed as (31) and (32).
The hyper-parameters θ = (τ01 2
, τ12
2
, τ13
2
, τ22
2
, τ23
2
, σ 2 , ρ) of the system noise
v0t , v kt and observation noise w t are the parameters which define the characteristic
of the prior distribution. To satisfy the conditions of the assumed model, we need to
decide the upper and lower limits for obtaining the maximum likelihood estimators
of the hyper-parameters by numerical nonlinear optimization (a quasi-Newton–
Raphson method). We set the initial conditions that the system noise variances
satisfy τ12 2
> τ22
2
, τ13
2
> τ23 2 1
.
1 The programming language of the numerical optimizer is Fortran 77.
PREDICTION OF INDIVIDUAL BOND PRICES 71

Figure 2. Comparison of AIC.

To estimate individual bond prices, we ran the first data sample period from
January 1980 to December 1989 and predicted their prices at the end of January
1990, after which we extended the data sample period monthly until November
2002. The results of the fitting for each type are summarized in terms of AIC values
at each time in Figure 2. The AIC values of the type 1 model are consistently
smaller than the ones of type 2, indicating that type 1 always fits better than type
2. Next the results of the out-of-sample forecasting of individual bond prices are
summarized in terms of the standard deviations of predicted price errors among
individual bonds in Figure 3, and the cumulative standard deviations of predicted
price errors in Figure 4. The standard deviation of predicted price errors is defined
by
 1/2
1
N
vt+1 = (Pi, t+1 (0) − E(Pi, t+1 (0)))2 . (40)
N i =1

The mean and distribution of standard deviations of predicted price errors are
summarized inTable I. From the graph and the table, it is observed that the prices
of the type 1 are better predicted than the prices of the type 2. Hence we limit our
attention mostly to the type 1. The estimated parameters of the type 1 are shown
in Figure 5. The parameters change over time. The changes of the coefficients δi jt
correspond to those of the term structures of the discount functions. There may be
some associations between the variations of δi jt ’s and certain fundamentals of the
economy. In fact, the coefficient δ13t is always negative and grows larger toward zero
after 1990, when the stock market bubble collapsed and the interest rates started to
72 HIROSHI TSUDA

Table I. Comparison of standard deviations of predicted price errors

Number in Each Size of Standard Deviation

Model Type Mean (yen) ∼2.0 ∼1.5 ∼1.0 ∼0.5 ∼0.0

Type 1 0.958 12 13 34 48 49
Type 2 1.262 31 13 33 45 34

Figure 3. Prediction standard deviations.

Figure 4. Cumulative prediction standard deviations.

PREDICTION OF INDIVIDUAL BOND PRICES 73

Figure 5. Graph of parameters (type 1).

74 HIROSHI TSUDA

Figure 6. Graph of hyper-parameters (type 1).

decline. The coefficient δ12t of coupon becomes almost zero after 1996, when the
interest rates remained ultra-low situation. On the other hand, the coefficients δ22t
and δ23t are almost positive. The coefficient δ22t of coupon moves together with the
fluctuation of interest rates, and becomes almost zero after 1996. The reason why
the coefficients δ12t ∼ δ23t move on toward zero is that the differences of individual
bond prices become smaller as the interest rates become lower. The estimates of
δ01t ’s have been mostly close 0.9 from 1980 to 1995 although they fluctuate greatly
in some months, but recently tend to decline. The hyper-parameters (τ01 2
, τ12
2
, τ13
2
,
τ22 , τ23 ) of the type 1 are shown in Figure 6. At each data sample period, the values
2 2

of the hyper-parameters are fixed.

PREDICTION OF INDIVIDUAL BOND PRICES 75

The estimated coefficients δi jt are apt to fluctuate. The reason for this phe-
nomenon is that the dimension of observation vector (N ) is much bigger than that
of state vector (smoothness prior). Kawasaki, Sato and Tachiki (1998) pointed out
the same phenomenon about multifactor model of equities. It is important to inves-
tigate how the predictive performance of the model for individual bond prices is
affected by making the estimated coefficients δi jt move smoothly.
Hence to make the estimated coefficients δi jt move smoothly, we tried to estimate
the time structures of the state vector x t by the following model:

x t =Fx t−1 + Gv t
y t =H t (x t + κt ) + w t , (41)

where the variance and covariance of system noise κt = (κ0t , κ1t , κ2t , κ3t , κ4t ,
0, . . . , 0) is expressed by
     ∗2 
κ0t 0 σ0 0 0 0 0
κ   0   0 σ ∗2 0  
 1t     1 0 0 
     ∗2 
 κ2t  ∼ N  0  , S =  0 0 σ2 0 0  . (42)
     ∗2 
 κ3t   0   0 0 0 σ3 0 
κ4t 0 0 0 0 0 σ4∗2
This state space model expresses the time structure of x t which consists of
signal and temporal additive noise. The state vector x t depends only on the noise
v t but not at all on κt . Kawasaki, Sato and Tachiki (1998) introduced a similar
model as temporal effect model for multifactor model of equities. This model can
be understood as a modeling of variance-covariance matrix in the observational
equation. This observational equation is equivalent to

y t = H t x t + w ∗t , w ∗t ∼ N (0, H t SH t + Rt ). (43)

We call this model a type 3 model. The results of the estimated coefficients δi jt of the
type 3 are shown in Figure 7. The upper and lower limits for obtaining the maximum
likelihood estimators of the hyper-parameters by the numerical optimization are
the same as in type 1. The estimated coefficients δi jt move smoothly as expected.
However the AIC values of this type are bigger than those of type 1, which is shown
in Figure 8. Hence the goodness of the fitting for this type becomes worse compared
with type 1.
Next the results of the prediction of individual bond prices are summarized in
terms of the standard deviations of predicted price errors among individual bonds
in Figure 9, and the cumulative standard deviations in Figure 10. From the graphs,
it is observed that the prices of the type 1 are better predicted than the prices of
type 3. Hence we found that the predictive performance of the type 1 model is best
among these models.
76 HIROSHI TSUDA

Figure 7. Graph of parameteres (type 3).

PREDICTION OF INDIVIDUAL BOND PRICES 77

Figure 8. The comparison of AIC.

Figure 9. Prediction standard deviations.

3.2. COMPARISON WITH AN APPROACH BASED ON THE TDM MODEL

Furthermore, we compared the predictive performance of the type 1 model with an

approach based on the TDM model. We cannot obtain the information on individual
future bond prices directly from the TDM model because a time series structure
is not included in the TDM model as we have already described. Accordingly, to
predict the parameters of the TDM model, we assumed a random walk model as
78 HIROSHI TSUDA

Figure 10. Cumulative prediction standard deviations.

same as the type 1 of our model:

 

δi jt = 
δi j,t−1 + i jt , i jt ∼ N 0, ςi2j . (44)

In other words, the expectations of parameters at time t + 1 become equivalent to

their values at time t.

E(
δi j,t+1 ) = 
δi jt . (45)

We call this approach a separate method. The results of the out-of-sample fore-
casting of individual bond prices are summarized in Table II in terms of standard
deviations of predicted price errors. We cannot judge which approach is superior
from the numbers shown in the table, because the differences in the predicted
price errors are small between two prediction procedures. The monthly differences
(our model–separate method) of the standard deviations of predicted price errors
are shown in Figure 11. We found that the predictive performance of our model
is somewhat superior to the separate method because the minus cases are more

Table II. Comparison of standard deviations of predicted price errors

Number in Each Size of Standard Deviation

Prediction Type Mean (yen) ∼ 2.0 ∼ 1.5 ∼ 1.0 ∼ 0.5 ∼ 0.0

Our model (type 1) 0.958 12 13 34 48 49

Separate method 0.972 11 13 38 46 48
PREDICTION OF INDIVIDUAL BOND PRICES 79

Figure 11. The monthly differences (our model–separate method) of the standard deviations.

frequent than the plus cases and larger in size. It is difficult to clarify the superiority
between two prediction procedures using predicted price errors because bond price
fluctuations are very small compared to those of stocks.
Consequently, the judgmental method that we adopted is to compare investment
performances of the portfolio consisting of more than one bond selected in the
order of high predicted return with equal weights. An investment performance of a
portfolio becomes higher as the models give us better prediction for a term structure
pattern of return. Figure 12 shows the accumulative profit of investing according to

Figure 12. The accumulative investment profit.

80 HIROSHI TSUDA

Figure 13. Term structure of predicted returns and realized returns.

a self-financing trading rule every month from January 1990 to December for 2002.
As the number of bonds of portfolios increase, the accumulative investment profit
of the portfolio gets closer to the profit of the portfolio consisting of all bonds with
equal weights. We found that the accumulative investment profit of the portfolio
consisting of the bonds chosen by our model was larger than the one of the separate
method.
Figure 13 shows the predicted returns and realized returns of individual bonds at
the end of August, 1998. The realized returns become bigger as a term to maturity
gets longer. Our model shows the same tendency for predicted returns, but in the
case of the separate method, it shows declining tendency in the part of long term to
maturity. The differences between the patterns of such a term structure of predicted
returns lead to a difference of accumulative investment profit. Figure 14 shows the
predicted prices for the type 1 of our model and the realized prices for individual
bonds at 1996.6 and 2001.2. Figure 15 shows the predicted returns and realized
returns of individual bonds simultaneously. In terms of returns, the patterns of the
term structure of the realized returns are well captured by those of the predicted
returns. This information for the term structure of the returns has practical benefits
for the construction of bond portfolios and bond trading.

4. Conclusions
In this paper, we presented a dynamic bond pricing model with a time series structure
which can obtain the information on individual future bond prices. The model has
the feature that the whole stochastic process of the random cash-flow discount func-
tions of each individual bond has a time series structure. The model prices individual
PREDICTION OF INDIVIDUAL BOND PRICES 81

Figure 14. Predicted prices and realized prices of individual bond.

bonds with different cash flows and different attributes. The model expresses the fea-
ture of correlation among individual bond prices by assuming a variance-covariance
structure for the random part of the stochastic discount function.
We tried to estimate and predict individual bond prices with this model. We
assumed three types of the time structures for the unobservable parameters. We
found that the type 1 model is the best and that the standard deviation of its predicted
price errors for the period 1990.1 through 2002.12 is 0.96 yen on the average,
where the face value of a JGB is around 100 yen. Among 156 (12 month × 13
years) months tested from 1990.1 through 2002.12, there were only 6 months with
standard deviations of predicted price errors greater than 3.0 yen in the case of the
82 HIROSHI TSUDA

Figure 15. Term structure of predicted returns and realized returns.

type 1 model. In fact, our model fits well in usual situation, but our model does
not fit properly when interest rates change drastically. This affects the predictive
performance of the model for prices. However, if such an event does not occur, the
patterns of the term structure of the realized returns are well captured by those of
the predicted returns of our model. Hence, using this information, we will be able
to make better investment decisions.

Acknowledgements
We thank chief editor and referees for useful comments on the previous version.
PREDICTION OF INDIVIDUAL BOND PRICES 83

Appendix
AKAIKE INFORMATION CRITERION (AIC)
Akaike information criterion (AIC)(Akaike, 1973, 1974) is one of the information
theoretic criteria for selecting the best of alternative parametric models based on
observed data. An approximate correction of the bias involved in the maximized
log-likelihood as an estimator of the average expected log-likelihood is considered
in AIC. The log-likelihood is an estimate of the information theoretic measure of
the dissimilarity between two distributions, and is called Kullback–Leibler infor-
mation. The approximate bias reflected in AIC is equal to the number of parameters
estimated in the model. A model with a smaller value of AIC is determined as a
better model.

KALMAN FILTER AND SMOOTHING

The state estimation and filtering and smoothing for the state space model are as
follows.

State Estimation and Identification

The problem of state estimation is to evaluate p(x t |x s ), the distribution of x t given
the observations Y s ≡ (y 1 , . . . , y m ) and the initial values x 0|0 and V 0|0 . The state
prediction, filtering and smoothing problems respectively refer to the situations in
which s < t, s = t, s > t.
The sequence of events in Kalman state space filtering is as follows. Starting with
the initial conditions x 0|0 and V 0|0 , the initial operation is to predict the next state
x 1 . In standard notation, the one-step-ahead predictor x 1|0 is computed. Following
that the observation y 1 is used in a filtering stage to compute the filtered estimate
of the state at time t = 1. That is represented by x 1|1 . The process is recursive,
prediction is the next operation and x 2|1 is computed followed by the filtering stage
during which x 2|2 is computed. The specific computations are given below by the
well-known Kalman filter formulas(e.g., Anderson and Moore, 1979).

Prediction (Time-Update)
From the filtered value x t−1|t−1 at time t − 1 and its covariance V t−1|t−1 , we obtain
the one-step-ahead predictor x t|t−1 at time t and its covariance V t|t−1 as follows:

x t|t−1 =Fx t−1|t−1

V t|t−1 =FV t−1|t−1 F  + GQG . (A1)

Here, Q is the covariance matrix for system noise V t . We assume that the initial
conditions x 0|0 and V 0|0 are given.
84 HIROSHI TSUDA

Filter (Measurement-Update)
Next, using the Kalman gain K t , state vector x t|t at time t and its covariance V t|t
are calculated by the following:

K t = V t|t−1 H t (H t V t|t−1 H t + Rt )−1

x t|t = x t|t−1 + K t (y t − H t x t|t−1 )
(A2)
V t|t= (I − K t H t )V t|t−1 .

Here, I is the unit matrix, and H t is the transpose of H t .

Smoothing
Using the outputs of the Kalman Filter, the smoothed state x t given the entire set
of observations y T is given by the fixed interval smoother:

At =V t|t F  V −1
t+1|t
x t|T =x t|t + At (x t+1|T − x t+1|t )
V t|T =V t|t + At (V t+1|T − V t+1|t )At . (A3)

References
Akaike, H. (1973) Information theory and an extension of the maximum likelihood principle. In
B. N. Petroc and F. Caski (eds.), Second Intern Symp in Information Theory, Akademiai Kiado,
Budapest, pp. 267–281.
Akaike, H. (1974) A new look at the statistical model identification. IEEE Trans. Auto. Contr., AC-19,
716–723.
Amin, K. I. and Morton. A. J. (1994) Implied volatility functions in arbitrage-free term structure
models. J. Financ. Econ. 35, 141–180.
Anderson, B. D. O. and Moore, J. B. (1979) Optimal Filtering, Prentice Hall, New Jersey.
Asano, H. and Takada, Y. (1989) The Estimation and the Application for the Term Structure of Interest
Rate. The Development of Modern Security Investment Method, Nihonkeizai (in Japanese).
Babs, S. H. and Nowman, K. B. (1999) Kalman filtering of generlized vasicek term structure models.
J. Financ. Quant. Anal. 34 (1), 115–130.
Bierwag, G. O. (1986) Duration analysis, Harper & Row, Publishers. Ballinger Publishing Co.
Brennan, M. J. and Schwartz, E. S. (1979) A continuous time approach to the pricing of bonds.
J. Banking and Financ. 3, 133–155.
Caines, P. E. and Rissanen, J. (1974) Maximum Likelihood estimation of parameters in multivariate
gaussian stochastic processes, IEEE Trans. Info. Theory, IT-20, 102–104.
Carlton, W. T. and Cooper, I. A. (1976) Estimation and uses of the term structure of interest rate.
J. Financ. 31, 1067–1083.
Chambers, D. R., Caleton, W. T. and Waldman, D. W. (1984) A new approach to estimation of the
term structure of interest rates. J. Financ. and Quant. Anal. 19, 233–252.
Coleman, T. S., Fisher, L. and Ibbotson, R. G. (1992) Estimating the term structure of interest rates
from data that include the prices of coupon bonds. J. Fixed Income, September, 85–116.
PREDICTION OF INDIVIDUAL BOND PRICES 85

Cox, J. C., Ingersoll, J. E. and Ross, S. A. (1981a) A Re-examination of traditional hypotheses about
the term structure of interest rates. J. Financ. 36, 769–799.
Cox, J. C., Ingersoll, J. E. and Ross, S. A. (1985) A theory of the term structure of interest rates.
Econometrica 53, 385–407.
Dothan, L. U. (1978) On the Term Structure of Interest Rates. J. Financ. Econ. 6, 59–69.
Edwin Elton, J. and Martin, J. Gruber. (1981) Modern Portfolio Theory and Investment Analysis,
Wiley, NewYork.
Geyer, A. and Pichler, S. (1999) A state-space approach to estimate and test multifactor Cox-Ingersoll–
Ross models of the term structure. J. Financ. Res. XXII, (1), 107–130.
Gultekin, N. B. and Rogaiski, R. J. (1984) Alternative duration specifications and the measurement
of basis risk: Empirical tests. J. Business 57, 2.
Heath, D., Jarrow, R. and Morton, A. (1990) Bond pricing and the term structure of interest rates: A
discrete time approximation. J. Financ. Quant. Anal. 25, 419–440.
Heath, D., Jarrow, R. and Morton, A. (1992) Bond pricing and the term structure of the interest rates;
A new methodology. Econometrica 60, 77–105.
Ho, Thomas S. Y. and Lee, S. B. (1986) Term structure movements and pricing interest rate contingent
claims. J. Financ. 41.
Houglet, M. X. (1980) Estimating the Term-Structure of Interest Rate for Nonhomogeneous Bonds,
Dissertation, Graduate School of Business, University of California, Berkeley.
Kariya, T. (1993) Quantitative Methods for Portfolio Analysis. Kluwer Academic Publishers.
Kariya, T. (1995) The Basis and Application of Bond Quantitative Analysis. Toyokeizai (in Japanese).
Kariya, T. and Tsuda, H. (1994) New Bond Pricing Models with Applications to Japanese Data.
Financ. Engineering and the Japanese Markets, Kluwer Academic Publishers, vol 1, 1–20.
Kawasaki, Y., Sato, S. and Tachiki, S. (1998) Smoothness Prior Approach to Estimate Large Scale
Multifactor Models. Research Memorandum, The Institute of Statistical Mathematics.
Kitagawa, G. and Gersch, W. (1984) A smoothness priors-state space modeling of time series with
trend and seasonality. J. Amer. Statist. Assoc. 79, 378–389.
Kitagawa, G. and Gersch, W. (1996) Smoothness Priors Analysis of Time Series. Springer.
Komine, M. (1989) The phenomena peculiar to Japanese bond market and the term structure analysis.
Financ. Review, October (in Japanese).
Litterman, R. and Scheinkman, J. (1991) Common factors affecting bond returns. J. Fixed Income,
June, 54–61.
McCulloch, J. M. (1971) Measuring the term structure of interest rates. J. Business, 19–31.
McCulloch, J. M. (1975) The tax adjusted yield curve. J. Financ. 30, 811–830.
Schaefer, S. M. (1981) Measuring a Tax Specific Term Structure of Interest Rtaes in the Market for
British Government Securities. Econ. Journal, 91, 415–438.
Simon Babbs, H. and Ben Nowman, K. (1999) Kalman filtering of generalized vasicek term structure
models. J. Financ. Quant. Anal. 34, 115–130.
Vasicek, O. A. (1977) An equilibrium characterization of the term structure. J. Financ. Econ. 5,
177–188.
Vasicek, O. A. and Fong, H. G. (1982) Term structure modeling using exponential splines. J. Financ.,
339–348.
Yamada, M. (1990) Bond price fluctuation analysis, The Basis and Application of Finance and Security
Quantitative Analysis, Toyokeizai.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.