A Principal-Component-Based Affine TSM
A Principal-Component-Based Affine TSM
A Principal-Component-Based Affine TSM
Structure Model
R. Rebonato, I Saroka, V Putyatin
PIMCO, Oxford University
June 16, 2014
Abstract
We present an essentially ane model with pricipal components as
state variables. We show that, once no-arbitrage is imposed, this choice
of state variables imposes some unexpected constraints on the reversionspeed matrix, whose N 2 elements can be uniquely specied by its N eigenvalues. The requirement that some of its elements should be negative gives
rise to a potentially complex dynamics, whose implications we discuss at
length. We show how the free parameters of the model can be determined
by combining cross-sectional information on bond prices with time-series
information about excess returns and by enforcing a smoothness requirement. The calibration in the P and Q measures does not require heavy
numerical search, and can be carried out almost fully with elementary
matrix operations. Once calibrated, the model recovers exactly the (discrete) yield cuirve shape, the yield covariance matrix, its eigenvalues and
eigenvectors. The ability to recover yield volatilities well makes it useful
for the estimation of convexity and term premia. The model also recovers
well quantities to which it has not been calibrated, and oers an estimation of the term premia for yields of dierent maturities which we discuss
in the last section.
The theory of ane and essentially ane models is well established. See, eg,
Bolder (2001) for a survey that covers both theory and implementation issues,
or Piazzesi (2010) for an up-to-date and comprehensive review.
Recently an interesting variation on this well-rehearsed theme has been introduced by Christensen, Diebold, Rudebush (2011), and developed in Diebold
and Rudebush (2013) who show how to turn the static (curve-tting) Nelson
and Siegel (1987) model into a dynamic ane model.1 To the extent that the
coecients of the Nelson-Siegel model generate a close match to the observed
term structure and it is well known that they do , the dynamic Nelson-Siegel
1 See
d
xt =
a (
x t ) dt + b (
x t ) d
zt
(1)
with
a0+
a 1
x t,
a (
x t) =
T
b ( x t) b ( x t)
= b0 + b1 x t ,
a0 RN , a1 RNN
b0 RNN , b1 RNNN
(2)
xt
rt = c0 +
c 1
(3)
then bond prices, PtT , can be written as exponentially ane functions of the
factors,
T
AT
t + Bt
PtT = e
xt
(4)
Following the notation in Dai and Singleton (2001), we focus in what follows
on models for which b1 = 0, in which case the factors follow an N -dimensional
Ornstein-Uhlenbeck process.
yT =
x
u +U
(5)
t
2 For
How the ane link between the factors and the yields is established provides a
useful classication perspective for exponentially ane models. More precisely,
in some approaches the factors are latent variables and the loadings (
u t and
U t ) are not specied a priori, (see, eg, DAmico et al (2004)), but are derived
from the no-arbitrage conditions and the calibration (eg, via Kalman ltering) of
the model. In other approaches, the modeller assigns a priori the link between
the factors and the yields. For instance, Due and Kan (1996) simply identify
process for the factors of pre-specied models, ie, when the loadings
u t and U t
are assigned a priori. In this paper we make use of these results for the special
case when the factors are chosen to be principal components.
In so doing we discover some interesting results: indeed, we show in Section
3 that it is possible to specify an innity of term structure models such that:
the driving factors are principal components;
they follow a men-reverting (generalized Ornstein Uhlenbeck) dynamics;
an arbitrary exogenous covariance matrix among N yields can always be
exactly recovered (and hence so are all the observed eigenvalues and eigenvectors);
an arbitrary exogenous yield curve (also dened by N yields) is exactly
recovered;
no-arbitrage is satised.
Accomplishing this, however, imposes some important constraints on the
mean-reverting dynamics, the reason for which is rather subtle. An intuitive
explanation of what these constraints entail goes along the following lines.
1.1
First of all, it is well known that, given an N -dimensional O-U process, diagonalizing via an orthogonal transformation either the diusion or the reversion5 Saroka
Most work in ane term-structure modelling straddles the physical and riskneutral measures. In one common approach (for a recent and popular example,
see, eg, DAmico et al (2004)), one starts from the estimation in the real-world
(P) measure of the statistical properties of some features of the state variables
(say, the reversion level)9 . In parallel, cross-sectional information about prices
allows the determination in the risk-neutral (Q) measure of the same riskadjusted statistical quantities. The market price of risk is then usually obtained
as the dierence (change of measure) between the two set of quantities. Which
7 tCherito, Filipovic and Kimmel (2010) shopw that, under loose conditions, it is possible
to diagonalize the diousion matrix using a regular, but not necessarily orthogonal, transformation see their Theorem 2.1 and Corollary 2.2.
8 as dened in Cherito, Filipovic and Kimmel (2010)
9 In this apprach, the P-measure statistical estimation is sometime supplemented by survey
data.
quantities require risk adjustment depends on the posited dependence (if any)
of the market price of risk on the state variables.
In a complementary approach (see, eg Cochrane and Piazzesi (2008)) one
links instead the two measures by looking at the excess returns produced by
systematically investing in long-dated (n-period) bonds and by nancing at the
1-period rate.
In our work we follow a variant of this approach. More precisely,
1. we start by determining the measure-invariant model parameters (the coecients of the diagonal diusion matrix) using real-world volatility data;
2. keeping these data xed, we determine the measure-dependent reversionspeed matrix in the Q measure by cross-sectional tting to the whole
covariance matrix;
3. wit this information, we carry out a cross-sectional t to the yield curve
to determine the reversion-level vector in the Q measure;
4. we then carry out an empirical study of excess returns, and we establish
(by multi-variate regression) a link between these excess returns and our
state variables;
5. as a next step, we determine (see Section 00) the shape of the dependence of the reversion-level vector and the reversion-speed matrix (in the
P measure) on the market prices of risk associated with our model and our
chosen state variables in order to accommodate the empirical ndings
in Duee (2002) and the results of our own studies, at this point we allow
for the market price of risk to depend in an ane manner on the state
variables, (ie, we require our model to be essentially ane);
6. nally, we specialize the results in point 5. above so as to reect the
particular dependence determined in our empirical estimation of excess
returns.
We stress that the last step is quite general, and does not rely on the specic empirical ndings of our statistical estimation. For instance, a Cochraneand-Piazzesi-like return-predicting factor (see Cochrane and Piazzesi (2005,
20008))10 or a slope factor (as in Duee, 2002) can be readily accommodated
by our methodology.
So, for the avoidance of doubt: we start from the Q measure and we determine
by cross-sectional t to bond prices the Q-measure model parameters; we carry
out a statistical estimate of excess returns; with this information we distil the
P-measure model parameters.
1 0 In order to accommodate exactly the Cochrane-Piazzesi tent factor, ve principal components would have to be used. Conceptually, our approach extends without diculty to as
many factors as desired. The uniqueness of parameters in the calibration phase may disappear
if too many factors are used.
3
3.1
The Set-Up
Notation
rt =
e T1
yt
3.2
(6)
Consider the following dynamics for the component yields of the N 1 vector
y:
dy = [...]dt + dwQ,P
(7)
with
= diag [1 , 2 , ..., n ]
E dwdwT = dt
(8)
(9)
and the drift term reecting the no-arbitrage conditions when dwQ,P = dwQ ,
E dy dy T = T = mkt dt
(10)
(11)
= diag [1 , 2 , ..., N ]
(12)
with
(13)
To the extent that the exogenous matrix mkt is positive denite, all the eigenvalues i are positive.
y t = y +
xt
(14)
rt =
e T1
yt =
e T1 y +
e 1
x
(15)
and therefore
3.3
e T1 y
0
(16)
T1
e T1
(17)
rt = 0 +
T1
x
(18)
d
xt = K
(19)
x t dt + S dz
and
E dz dz T = Idt
(20)
(21)
and we impose
si =
(22)
For the reasons discussed in the introductory section, we would also like
the reversion-speed matrix, K, to be diagonal, but, at this stage, we do not
know whether this is possible (once a diagonal form is imposed for the diusion
matrix) indeed, we shall see that it is not.
1 1 Saroka
T
t
rs ds
(23)
PtT = E Q e
3.4
0 +
T
(
1 x s )ds
(24)
Solution
It is well known12 (see, eg, Dai and Singleton (2000)) that the solution to Equation (24) is given by
PtT = exp
T
AT
t + Bt
xt
(25)
T
T
with the vector B t and the scalar At satisfying the ordinary dierential equations (with = T t)
1
T
T
dA
= 0 + B
K +
B
SS T B
d
2
dB
=
T1 K B
d
with boundary conditions
B ( = 0) = 0, A ( = 0) = 0
B =
eK
e 1 d
(26)
(27)
(28)
(29)
(30)
with
K = diag [lj ] ,
j = 1, 2, .., N
(31)
1
B = adiag
1 elj
lj
1 2 All
a1 T
e1
(32)
Once the vector B has been obtained, the scalar A can be obtained as
A =
0
0 + B
K +
B
2
SS T B d
(33)
In the case we consider, the integral can be carried out analytically, and the
resulting expression is given in Appendix A.
3.5
vector,
y t , has the expression
y =
x
(34)
t
with
A 1
1
A 2
2
=
...
A N
N
and
B 1
B 2
=
2
...
B N
N
(35)
(36)
At the same time, for identiability of the factors with principal components,
Equation (14) must also hold:
y t = y +
xt
(37)
(38)
y =
(39)
A (+0)
+0
A ( k )
=
,
k
=
(40)
k = 2, 3, ..., N
(41)
In sum: if the vector y is chosen as per Equations (40) to (41), the time-0
discrete yield curve is automatically and exactly recovered for any reversion
level vector, . (We discuss in the calibration section how a good choice the
be tackled clearly shows at least one advantage from identifying the vector
xt
with the principal components. We therefore turn in the next section to the
showing that the identication is indeed possible.
Before that, we note in passing that the rst element of the vector y is at
this point indeterminate, ie, any value can be chosen for it, and all N yields can
be recovered exactly.13 This can be seen as follows. Recall that the bond price
is given by
PtT = exp
T
AT
t + Bt
Ai
i
xt
(42)
(43)
and y =
(with Ai ATt i ). Therefore Ai = i yi , and A1 = 1 y1 in
particular. We know, however, that
A ()
B ()
lim
0
lim
= 0
(44)
= 1
(45)
and therefore we see from Equation (40) that any value can be assigned to y1 ,
while retaining the property that the innitesimally short yield be given by
lim
(T t)0
4
4.1
ytT = rt
(46)
Results
Impossibility of Identication When K Is Diagonal
10
To prove that the reversion-speed matrix K cannot be diagonal when the state
variables are principal components, we proceed by reduction ad absurdum, ie,
we show that, given Equation (47), an impossibility arises.
So, lets assume that K is diagonal and the state variables independent. In
this setting it is straightforward to show that the matrix is given by14
[ 11 , 12 , ..., 1N ]
22 2
N N 2
1 1e11 2
11 , 1e22
12 , ..., 1eNN
1N
2
11
=
(48)
...
11 N
22 N
NN N
1
1e
11 , 1e 22
12 , ..., 1e NN
1N
N
11
with
e T1
[11 , 12 , ..., 1N ] =
(49)
ie, the row vector whose elements are the rst row of the eigenvector matrix,
. (The rows 2 to N are straightforward. The rst row is obtained by recalling
from Equation (45) that the limit of B () / as goes to zero.)
Consider now T . The element T
is indeed equal to 1 (as it should
11
r1
= r1
(50)
1 ejj s 2
1j
jj
(51)
In reality we have:
r1
1
r
But this term cannot possibly be zero for r = s, because we know that
21j =
jj s
4.2
11
1eii 2
ii
is simply a
(53)
with the eigenvalues {lj } distinct and real. Let F be the [N N ] matrix of
elements [F ]ij given by
1 1 eli j
Fij
.
(54)
j
li
Then for any 2 , 3 , .., N and for any l1 , l2 , ..., lN such that all the distinct
and real eigenvalues are also positive (so as to ensure stability of the dynamical
such that
K = T F 1 K F
(55)
T = I
(56)
4.3
12
(32) and (36), that play a central role in determining the prices and yields of
bonds, and in ensuring the orthogonality. But the rotation of axes imposed
by the principal-component interpretation of the factors also aects, in a less
obvious way, the admissible reversion-speed matrices, which, we recall, are given
by
K = T F 1 K F
(57)
This link between the reversion speed matrix and the particular rotation singled out by entails a rather complex mean-reverting deterministic dynamics.
This can be see as follows.
Once the state variables have been chosen to be principal components, and
they have been assigned a mean-reverting behaviour as in Equation (19), no
13
y0T =
y0T E Q
1
T
rs ds
(59)
Therefore, simply by averaging out to a horizon T the values of the short rate
along a deterministic path, one can immediately relate the path of the short
rate to the current model yield curve. At the reference points, by construction,
one will observe a match between the market and the model yields (again,
within the limits of the approximation above); in between the reference points,
however, every choice of eigenvalues will determine, by aecting the path of the
state variables to their reversion levels, the values of the intermediate-maturity
yields. The observed market yields therefore behave as xed knots through
which the yield curve has to move: how smoothly it goes from point to point
will depend on the eigenvalues of the reversion speed matrix. This is clearly
shown in Fig (1):
The three lines in Fig (1) show the averages of the short rate out to ve
years for the eigenvalues of the reversion-speed matrix used in the case study
(curve labelled Base); for eigenvalues twice as large (curve labelled Base*2);
and for eigenvalues four times as large (curve labelled Base*4). In all cases,
to within the accuracy of the approximation, the average of the short rate out
to ve years is indeed 2.00% (the exogenously assigned market value of y02 );
however, intermediate yields can assume values which strongly depend on the
eigenvalues of the reversion-speed matrix. The larger these eigenvalues, the
more complex the behaviour in between the knots.
This interpretation also makes more precise the concept of complexity for
the yield behaviour, to which we have frequently alluded to above: for instance,
as in the case of splines, the integral of the non-convexity-induced curvature of
the model yield curve between reference points can be taken as a measure of
complexity.
Similar considerations apply to the interpolated covariance matrix. Also in
this case, the choice of the eigenvalues l strongly inuences the values of the
interpolated covariance elements. Indeed, some choices for the vector l can
even produce negative correlations for yields in between the exactly recovered
covariances.
Figure (2) shows the square root of the entries of the model (top panel) and
market (bottom panel) covariance matrix for yields from 1 to 30 years obtained
with the optimal choice of the eigenvalues l .18 The overall quality of the t
1 8 Unless
otherwise stated, in all our calibration studies we used N = 3 , ie, 3 yields, and 3
14
Figure 1: The averages of the short rate out to ve years for the three cases
discussed in the text, namely, for the eigenvalues of the reversion-speed matrix
used in the case study (curve labelled Base); for eigenvalues twice as large
(curve labelled Base*2); and for eigenvalues four times as large (curve labelled
Base*4). In all cases, to within the accuracy of the approximation, the average
of the short rate out to ve years is indeed 2.00% (the exogenously assigned
market today value of the ve-year yield).
15
for the inter- and extrapolated covariance matrix is excellent, with a maximum
error of 6 basis points (in units equivalent to absolute volatility) and an average
absolute error of 1.5 basis points (in the same units).
Figure 2: The square root of the entries of the model (top panel) and market
(bottom panel) covariance matrix for yields from 1 to 30 years for the optimal
choice of the eigenvalues l . (Note that the intervals along the x and y axes
are not equally spaced units in years.)
One might think that, since the values of the covariance matrix in correspondence with the reference yields are exactly recovered by construction, the
errors in interpolation (and possibly extrapolation) should be small. If this were
true, little information about the eigenvalues l could be gleaned from the
principal components.
16
Figure 3: Same above for a poor choice of the eigenvalues l . Note that
the covariance matrix elements between each reference yield are still exactly
recovered.
This dependence of the intermediate yields and of the intermediate co
17
We want to show in this section how the model behaviour can be specied both
in the Q (risk-neutral) and in the P (data-generating) measures.
Lets go back to the Q-measure dynamics (19) which we re-write for ease
of reference:
x t dt + S dz Q
d
xt =K
(60)
As discussed in Section 2, we are going to assign to the market price of risk,
T t , one of the ane forms discussed in the literature, nested in the following
general formulation:
q 0 +R
(61)
T t =
xt
For instance, if we embraced the Duee (2002) specication (according to which
the magnitude of the market price of risk depends on the slope of the yield curve)
we would have19 for the matrix R
0 a 0
R= 0 b 0
(62)
0 c 0
In general, for any specication of the dependence of the market price of risk
on the state variables, we have
d
xt =K
q 0 +Rt
x t dt + S dz P + S (
x t ) dt
(63)
x t dt + S dz P
(64)
d
x t = KP
with
KP = K SR
and
P
1
= (K SRt )
K + S
q0
(65)
(66)
Equations (65) and (66) dene the reversion-speed matrix and the reversion-level
vector, respectively, as a function of the corresponding Q-measure quantities,
18
Estimating the Parameters of
q 0 and R
Using fty years of data from the data base provided by the Federal Reserve
Board of Washington, DC, (Gurkaynak et al, 2006), we have statistically estimated the excess returns from holding bonds up to 10 years. We have regressed
these excess returns against our state variables, ie, the principal components.20
the vector of excess returns, the OLS estimation gives
If we call
rx
rx
a + b
xt
t
(67)
E [rxt ] dt = E
rt dt = Dur S [
q 0 + R
x t ] dt
PtT
where
[Dur]ij =
1 PtTi
PtTi xj
(68)
(69)
a = Dur S
q0
(70)
b = Dur S R
(71)
q 0 = (Dur S)1
a
(72)
R = (Dur S)1 b
(73)
log PtTi
1 PtTi
xj
PtTi xj
(74)
and
and therefore
Recalling that
PtT = exp
we have
and therefore
T
AT
t + Bt
xt
Dur = (B t )T
q 0 = (B t )T S
R = (B t )T S
(75)
(76)
(77)
(78)
2 0 The method used is presented in a separate paper. The results we use here are independent
of the specic ndings.
19
Note that the duration matrix (B Tt ) clearly depends on the maturity of the
yield under consideration; so does the matrix of regression coecient, b, and the
vector of intercepts,
a . However, the market price of risk must be independent
eigenvalues of the reversion-speed matrix), the vector , and the rst element
7
7.1
7.2
equate these quantities to the reversion levels, P , in the P measure; to translate this vector to the Q measure using Equation (66) (which contains the
20
section. We also choose a trial value for the Q-measure reversion-level vector,
Q
.
Given these quantities, and after estimating the market price of risk vector,
T =
q 0 +Rt
x t , we can evolve in the P measure the state variables or the yields
from today to any horizon, , using the evolution equations
d
x t = KP
x t dt + S dz P
(79)
d
yt=
KPy
P
y t dt + S dz P
y
(80)
with
KPy = KQ SR T
and
P
y = KQ SR
= Q + I KQ
(81)
SR T y + KQ
S
q0
(82)
This P-measure projection can be carried out exactly for any horizon, . In
particular, it can be carried out for = . For this projection horizon, the
= E P [
y = KQ SR
y ]
(83)
Note that these model quantities, which have been obtained after a P
Q
on the vector .
Separately from this, we can also calculate from our historical record the
P
As a last step, we can equate the measured, y , and the model, E [ y ],
quantities
y = E P [
y ] = KQ SR
=
KQ SR
Q
+ I KQ
SR T y + KQ
S
q0
(84)
We can therefore determine the quantity, Q , that best achieves this match
between the model-implied and the empirically-observed long-term yields. (Note
that the reversion-speed vector, Q , also enters the expression for y . The
y depends on Q linearly.)
21
averages,
y , one can rest assured that, by using these reversion levels, the
7.3
trized by the arbitrary value of the rst element, y1 , of the vector, y . More
a yield longer than the maximum maturity yield in the set of N yields {
y }.
With this last piece of information the model is fully calibrated.
Results
We show in this section the result obtained by calibrating the model using the
procedure described above to Treasury data provided by the Fed (Gurkaynak et
al, 2006) for the period 1990-2014. Qualitatively similar results were obtained
using a longer data sets (extending back to the late 1970s. However we used the
shorter time window of 24 years in order to avail ourselves of information about
the 20 year yields, useful for assessing the quality of the extrapolation. We used
5 and 10 years for the intermediate and long yield maturities, respectively.
After using covariance information from the period 1990 to 2014, the eigenvalues of the Q-measure reversion-speed matrix, KQ , turned out to be
Eigenvalues of KQ
l1 = 0.03660
(85)
l2 = 0.63040
l3 = 0.63036
-0.1880 1.0187
-3.0499
KQ = -0.2229 0.7731
-4.2299
0.0242
-0.0304 0.7123
(86)
As anticipated, note the presence of large and negative reversion speeds, also
on the main diagonal.
With this reversion-speed matrix the model covariance matrices, and the
dierence between the model and the empirical covariance matrices were calculated. The results are shown in Fig (4) below:
22
Figure 4: Model and market covariance matrices, and the error (market - model).
The rows and columns correspond to maturities from 1 to 10 years.
23
The quality of the t is evident. To assess the ability of the model to predict
quantities it has not been calibrated to, the empirical and model yield volatilities
out to 20 years are shown in Fig (5). We stress that the volatilities beyond 10
years are extrapolated by the model.Fig (6) shows the observed and tted yield
Figure 5: Model and empirical yield volatilities. Volatilities beyond 10 years are
extrapolated.
curves on randomly selected dates between 1990 and 2014. The same model
parameters were used for all the ts, and only the state variables were changed.
They yields for the key maturities are, of course, perfectly recovered. It is
interesting to note, however, that the intermediate and the extrapolated yields
(ie, the yield beyond 10 years) are also well recovered.
After empirical estimation of the regression matrix of excess returns we were
able to estimate the P-measure reversion levels, and the trace of the matrix. See
Equations (65) and (66). To show the robustness of the procedure, we present
the estimates obtained for three dierent subsections of the data. Observe that
all the parameters remain reasonably stable, with the possible exception of the
reversion level of the rst factor in the rst half of the sample. This is probably
due to the quasi-unit-root nature of the rst principal component.
Data sample P1
P2
P3
Trace KP
1990-2002
(87)
2002-2014
Figure 6: Fitted and empirical yiled curves for random dates between 1990
and 2014. Note that the same parameters are used throughout, and that yield
beyond 10 years are extrapolated.
25
Once the term premia have been estimated, we can calculate the deterministic evolution of the reference yields under both measures from todays yield
curve (30-Jan-2014). This is shown in Fig (??).
of making use of the full covariance matrix information in the tting phase.
Figure 8: Yields and value of convexity for dierent eigenvalues of the meanreversion speed matrix K. Solid lines represent the yield curves, and dashed lines
represent the yield curves without the convexity term. The colours correspond
to dierent eigenvalues of K: blue - (0:02; 0:2; 0:5), green - (0:03; 0:3; 0:6), red
- (0:04; 0:4; 0:7).
Finally we show in Fig (9) the time series of the 10-year yield observed in
the market (yield under Q), and the yield that would have been observed if
investors had the same expectations, but were risk-neutral (yield under P). In
the same gure we also show the term premium (red line) which is the dierence
between the two yields. The average risk premium for the last 24 years averages
around 3%, which compares well with our empirical estimates of unconditional
excess returns for the 10 year maturity, shown in Fig (??).
Finally, we show in Figs (10) and (11) the model and empirical asymptotic
autocorrelation for the three principal components. Empirically, it is well-known
that the rst principal component should be far more persistent than the second
and the third (see, eg, the discussion in Diebold and Rudebush (2013).) This
is not borne out by our model, which also gets the overall speed of decrease of
the autocorrelation signicantly wrong.
27
Figure 9: Observed 10-year zero coupon yield (blue), the P10-year zero-coupon
yield (green) and the term premium (red).
28
29
Figure 11: Estimated empirical asymptotic autocorrelation of principal components. Observe that empirically the rst PC features a much higher autocorrelation than what the model produces. The model is not able to capture this
feature.
30
Conclusions
We have presented the theoretical results for the ane evolution of meanreverting principal component models, we have shown how the model can be
eectively calibrated using both Q- and P-measure information.
We have stressed that the no-arbitrage constraints imposed by using a prespecied model, and the choice of principal components as the mapping between
the model state variables and the observable yields aect deeply the model
evolution. In particular, we have shown that the reversion speed-matrix must
contain negative and large entries. This contributes strongly to the models
dynamic richness and complexity.
Once the calibration has been carried out, we have presented the model
behaviour.
With the exception of the degree of persistence of the principal components,
the behaviour of the model seems to be robust, believable and interpretable.
The cross-measure results of course depend strongly on the estimation of the
state dependence of the market price of risk (the excess returns). For illustrative
purposes, we have carried out in this study a rather simple-minded, regressionbased estimate. There is plenty of room for more careful analysis here (and the
Cochrane tent-shaped return-predicting factor could be accommodated, if one
so wished, in the modelling framework). We are planning to present the results
of our empirical study of excess returns for Treasuries as a separate piece of
work.
Acknowledgement 3 It is a pleasure to acknowledge useful suggestions provided by Dr Naik Vasant and Dr Andrei Lyashenko.
31
References
32
33
10
(88)
1
c = U I + A + A2 + ... U T
2
(89)
1
c = U IU T + U AU T + U A2 U T + ... =
2
(90)
1
I + UAU T + UA2 U T + ...
(91)
2
where the last line follows because of the orthogonality of U . Consider now the
T
exponential d = eU AU :
T
d = eUAU = I + U AU T +
1
U AU T
2
+ ... =
1
I + U AU T + U AU T U AU T + ... =
2
1
I + U AU T + U AAU T + ...
2
1
I + UAU T + UA2 U T + ...
(92)
2
Repeating for higher-order terms and comparing Equations (92) and (91), we
can conclude
T
U eA U T = eU AU
(93)
The same proof applies to non-orthogonal matrices, as long as matrix inversion
replaces transposition:
1
ZeA Z 1 = eZAZ
(94)
11
Let the solution for the bond price in a multidimensional OU process be given
by
Pt = eA( )+ B ( ) x t
(95)
34
with
Pt k
(96)
xi
(Note the positive sign of the exponent and the transpose.) Given the process
Bik () =
d
xt =K
x dt + S dz
d B ( )
= B ( ) T
e1
d
B (0) = 0
(97)
(98)
with
T = K
This is a inhomogeneous ODE. We solve it by i) nding the solution to the
homogeneous ode; ii) nding a particular solution; iii) joining the two and iv)
satisfying the boundary condition.
An aside: We have the transpose of K in Equation (97) because the expres
d B T ( )
e T1
(99)
= K B T ( )
d
and therefore
11.1
d B T ()
= B T K ( )
e T1
d
d B ( )
e1
= B ( ) T
d
=
(100)
d B hom ( )
+ B hom ( ) = 0
d
By inspection, an obvious candidate solution is
B hom ( ) = e H
Indeed
(101)
(102)
d B hom ()
= e H
d
(103)
d B hom ()
+ B hom ( ) = e H + e H = 0
d
(104)
and therefore
35
11.2
B () = C
(105)
d B ()
dC
=
= 0
d
d
and therefore
11.3
0 = C T
e1
1 T
C = e 1
(106)
(107)
(108)
We have found
B ( ) = e H + C = e H 1 T
e1
(109)
B (0) = e0 H 1 T
e1= 0
(110)
H = 1 T
e1
(111)
and therefore
Therefore we have
B () = e 1 T
e 1 1 T
e1 =
e I 1 T
e1
11.4
(112)
e( s) ds =
Int =
e( s) ds
(113)
1 = e I 1 T
e1
(114)
1 = e( s)
B ( ) =
e( s) T
e 1 ds
36
(115)
11.5
T
B =
e I 1 T
e1
e T1 1
(116)
T
B =
e( s) T
e 1 ds
ds =
e T1 e( s)
12
e( s) T
e1
ds
(117)
We obtained that
B ( ) =
e( s) T
e 1 ds
(118)
(119)
This gives
B ( ) =
e( s) T
e 1 ds =
ea a
( s)
T
e 1 ds =
= a
a
with
(120)
e ( s) ds a1 T
e1=
e ( s) ds a1 T
e1
Now,
e ( s) ds =
0
37
(121)
1el1
l1
0
1el2
l2
0
0
0
0
0
0
0
1el3
= D ( ) = diag
and, nally,
13
0
0
l3
...
...
...
...
...
...
1elN1
lN1
1 elii
,
lii
i = 1, N
0
0
0
0
0
1elN
lN
(122)
e1
B ( ) = a = aD ( ) a1 T
(123)
We obtained that
B ( ) = a = aD ( ) a1 T
e1
with :
and
= a a1
(124)
D () a1 T
e1
(125)
lii
1e
lii
D () = diag
i = 1, N
(126)
dA
= y1 + B T K + B T SS T B
d
2
(127)
(128)
We have to solve
Therefore
with
Int1 =
y1 ds = y1
(129)
T
B () ds K
(130)
B ( ) SS T B ( ) ds
(131)
Int2 =
0
Int3 =
1
2
38
Evaluation of Int3
13.1
We have
Int3 =
B ( ) SS T B () ds =
SS T a ds =
T T T
a SS a ds =
Q ds
(132)
Q = aT SS T a
(133)
f = a1 T
e1
(134)
= D f
(135)
with
To lighten notation dene
then
and therefore
Int3 =
0
T
f
Q ds =
DT QDds f
(136)
14
Let
d
xt = K
x t dt + S dz
(137)
P (xt |x0 ) = N (
t , t )
(138)
T = eKt + I eKt
(139)
Then
with
GT
(140)
39
du
ij
(141)
Then we have
[]ij = G1
T
0
(142)
ij
(143)
Then we have
and therefore
1 T
[]ij = G
e(i +j )u hij du G1
= G1
with
M = hij
15
MG1
1 e(i +j )u
i + j
(144)
(145)
(146)
We have dened
B 1
B 2
=
2
...
B N
N
B ( ) =
(147)
eK( s) T
e 1 ds
(148)
and that
T =
(149)
T = I
(150)
Therefore
I = T = = diag
1
1
1
0
T
1
eK( 1 s) ds
e 1 , ...,
N
40
N
0
1
B=
eK( N s) ds
e1
(151)
Dene
C = KT
(152)
C = T T 1
(153)
U = T 1
(154)
Orthogonalize C:
and dene
Then we have
I=
1
1
!
1
0
T e( 1 s) T 1 ds
e 1 , ...,
N
1 e
T diag
i 1
T e( N s) T 1 ds
e1
0
i
1 e N
i N
e 1 , ..., T diag
T 1
e1
T 1
"
=
(155)
with
1 eC 1
1 eC N
diag
U
e 1 , ..., diag
i
1
i N
u11 f11
u21 f21
U =
...
uN1 fN1
u11 f12
u21 f22
...
uN1 fN2
...
...
...
U
e1
u11 f1N
uN1 f2N
...
uN1 fNN
"
(156)
(157)
fij =
1 eC j
i j
(158)
(159)
U = DF
(160)
(161)
with
D = diag [u11 , u21 , ..., uN1 ]
u1
i1
and
are chosen in such a way that the rst, second, ..., nth column of F
are normalized. Therefore the matrix F consists of indepedent vectors in its
columns, and u1
i1 normalize it to a basis.
Putting the pieces together one gets
C = KT
41
(162)
and therefore
K = T C = T T T 1 =
T F 1 F
(163)
(164)
(165)
K =
(166)
and
42