Price of a ˜standard house™ from 1980 to 2000
0.80
0.60
0.40
0.20
0.00
0.20
1980:1 1985:1 1990:1 1995:1 1999:4
Figure 13.3. Smoothed common price component. Con¬dence intervals
are calculated for the 90% level.
XFGsssm6.xpl
[1,] "==========================================="
[2,] " Estimated hedonic coefficients "
[3,] "==========================================="
[4,] " Variable coeff. tStat. pvalue "
[5,] "  "
[6,] " log lot size 0.2664 21.59 0.0000 "
[7,] " log floor area 0.4690 34.33 0.0000 "
[8,] " age 0.0061 29.43 0.0000 "
[9,] "==========================================="
Table 13.4. Estimated hedonic coe¬cients β. XFGsssm6.xpl
zero. Thus, it is not surprising that the Kalman smoother produces constant
estimates through time for these coe¬cients. In the Appendix 13.6.2 we give
a formal proof of this intuitive result.
302 13 A simple state space model of house prices
The estimated coe¬cient of log lot size implies that, as expected, the size of the
lot has an positive in¬‚uence on the price. The estimated relative price increase
for an one percent increase in the lot size is about 0.27%. The estimated e¬ect
of an increase in the ¬‚oor space is even larger. Here, a one percent increase in
the ¬‚oor space lets the price soar by about 0.48%. Finally, note that the price
of a houses is estimated to decrease with age.
13.6 Appendix
13.6.1 Procedure equivalence
We show that our treatment of missing values delivers the same results as the
procedure proposed by Shumway and Sto¬er (1982; 2000). For this task, let us
assume that the (N — 1) vector of observations t
yt = y1,t . y3,t . y5,t ... yN,t
has missing values. Here, observations 2 and 4 are missing. Thus, we have only
Nt < N observations. For Kalman ¬ltering in XploRe, all missing values in yt
and the corresponding rows and columns in the measurement matrices dt , Zt ,
and Ht , are deleted. Thus, the adjusted vector of observations is
yt,1 = y1,t y3,t y5,t ... yN,t
where the subscript 1 indicates that this is the vector of observations used in the
XploRe routines. The procedure of Shumway and Sto¬er instead rearranges the
vectors in such a way that the ¬rst Nt entries are the observations”and thus
given by yt,1 ”and the last (N ’ Nt ) entries are the missing values. However,
all missing values must be replaced with zeros.
For our proof, we use the following generalized formulation of the measurement
equation
µm
yt,1 d Z
= t,1 + t,1 ±t + t,1
µm
yt,2 dt,2 Zt,2 t,2
and
µm Ht,11 Ht,12
cov t,1 = .
µm Ht,12 Ht,22
t,2
yt,1 contains the observations and yt,2 the missing values. The procedure of
Shumway and Sto¬er employs the generalized formulation given above and sets
13.6 Appendix 303
yt,2 = 0, dt,2 = 0, Zt,2 = 0, and Ht,12 = 0 (Shumway and Sto¬er, 2000, p. 330).
We should remark that the dimensions of these matrices also depend on t via
(N ’Nt ). However, keep notation simple we do not make this time dependency
explicit. It is important to mention that matrices with subscript 1 and 11 are
equivalent to the adjusted matrices of XploRe™s ¬ltering routines.
First, we show by induction that both procedures deliver the same results for
the Kalman ¬lter. Once this equivalence is established, we can conclude that
the smoother also delivers identical results.
PROOF:
Given µ and Σ, the terms a10 and P10 are the same for both procedures. This
follows from the simple fact that the ¬rst two steps of the Kalman ¬lter do not
depend on the vector of observations (see Subsection 13.5.3).
Now, given att’1 and Ptt’1 , we have to show that also the ¬lter recursions
at = att’1 + Ptt’1 Zt Ft’1 vt , Pt = Ptt’1 ’ Ptt’1 Zt Ft’1 Zt Ptt’1 (13.13)
deliver the same results. Using ss to label the results of the Shumway and
Sto¬er procedure, we obtain by using
Zt,1
def
Zt,ss =
0
that
Zt,1 Ptt’1 Zt,1 0 Ht,11 0
Ft,ss = + .
0 Ht,22
0 0
The inverse is given by (Syds¦ter, Strøm and Berck, 2000, 19.49)
’1
Ft,1 0
’1
Ft,ss = (13.14)
’1
0 Ht,22
where Ft,1 is just the covariance matrix of the innovations of XploRe™s proce
dure. With (13.14) we obtain that
’1
’1
Zt,ss Ft,ss = Zt,1 Ft,1 0
and accordingly for the innovations
vt,1
vt,ss = .
0
304 13 A simple state space model of house prices
We obtain immediately
’1 ’1
Zt,ss Ft,ss vt,ss = Zt,1 Ft,1 vt,1 .
Plugging this expression into (13.13)”taking into account that att’1 and
Ptt’1 are identical”delivers
at,ss = at,1 and Pt,ss = Pt,1 .
This completes the ¬rst part of our proof.
The Kalman smoother recursions use only system matrices that are the same
for both procedures. In addition to the system matrices, the output of the
¬lter is used as an input, see Subsection 13.5.5. But we have already shown
that the ¬lter output is identical. Thus the results of the smoother are the
same for both procedures as well.
13.6.2 Smoothed constant state variables
We want to show that the Kalman smoother produces constant estimates
through time for all state variables that are constant by de¬nition. To proof
this result, we use some of the smoother recursions given in Subsection 13.5.5.
First of all, we rearrange the state vector such that the last k K variables
are constant. This allows the following partition of the transition matrix
T11,t+1 T12,t+1
Tt+1 = (13.15)
0 I
with the k—k identity matrix I. Furthermore, we de¬ne with the same partition
˜ ˜
P P12,t
˜ def
Pt = Tt+1 Pt Tt+1 = ˜11,t ˜
P12,t P22,t
The ¬lter recursion for the covariance matrix are given as
Pt+1t = Tt+1 Pt Tt+1 + Rt+1
where the upper left part of Rt+1 contains the covariance matrix of the dis
turbances for the stochastic state variables. We see immediately that only the
˜
upper left part of Pt+1T is di¬erent from Pt .
13.6 Appendix 305
Our goal is to show that for the recursions of the smoother holds
M11,t M12,t
Pt— = , (13.16)
0 I
where both M s stand for some complicated matrices. With this result at hand,
we obtain immediately
ak = ak k
t+1T = aT (13.17)
tT
for all t, where ak contains the last k elements of the smoothed state atT .
tT
Furthermore, it is possible to show with the same result that the lower right
partition of PtT is equal to the lower right partition of PT for all t. This lower
right partition is just the covariance matrix of ak . Just write the smoother
tT
recursion
PtT = Pt (I ’ Tt+1 Pt— ) + Pt— Pt+1T Pt— .
Then check with (13.15) and (13.16) that the lowerright partition of the ¬rst
matrix on the right hand side is a k—k matrix of zeros. The lowerright partition
of the second matrix is given by the the lowerright partition of Pt+1T .
PROOF:
Now we derive (13.16): We assume that the inverse of Tt+1 and T11,t+1 exist.
The inverses for our model exist because we assume that φ2 = 0. For the
partitioned transition matrix (Syds¦ter, Strøm and Berck, 2000, 19.48) we
derive
’1 ’1
T11,t+1 ’T11,t+1 T12,t+1
’1
Tt+1 = . (13.18)
0 I
Now, it is easy to see that
’1 ˜ ’1
Pt— = Tt+1 Pt Pt+1t . (13.19)
We have (Syds¦ter, Strøm and Berck, 2000, 19.49)
˜ ’1
˜
’∆t P12,t P22,t
∆t
’1
Pt+1t = (13.20)
’1 ˜ ˜ ’1 + P ’1 P12,t ∆t P12,t P ’1
˜ ˜ ˜ ˜ ˜
’P22,t P12,t ∆t P22,t 22,t 22,t
with ∆t as a known function of the partial matrices. If we multiply this matrix
˜
with the lower partition of Pt we obtain immediately [0 I]. With this result
and (13.18) we derive (13.16).
306 13 A simple state space model of house prices
Bibliography
Bailey, M. J., Muth, R. F. and Nourse, H.O. (1963). A regression method for
real estate price index construction, Journal of the American Statistical
Association 58: 933“942.
Baumol, W. (1959). Economic Dynamics, 2nd ed., Macmillan, New York.
Bera, A. K. and Jarque, C. M. (1982). Model Speci¬cation Tests: a Simulta
neous Approach, Journal of Econometrics 20: 59“82.
Cho, M. (1996). House price dynamics: a survey of theoretical and empirical
issues, Journal of Housing Research 7:2: 145“172.
Clapp, J. M. and Giaccotto, C. (1998). Price indices based on the hedonic
repeatsales method: application to the housing market, Journal of Real
Estate Finance and Economics 16:1: 5“26.
Durbin, J. and Koopman, J. S. (2001). Time Series Analysis by State Space
Methods, Oxford University Press, Oxford.
Engle, R. F. and M. W. Watson (1981). A OneFactor Multivariate Time Series
Model of Metropolitan Wage Rates, Journal of the American Statistical
Association 76: 774“781.
Gourieroux, C. and Monfort, A. (1997). Time Series and Dynamic Models,
Cambridge University Press, Cambridge.
Greene, W. H. (2000). Econometric Analysis. Fourth Edition, Prentice Hall,
Upper Saddle River, New Jersey.
Hamilton, J. D. (1994). Time Series Analysis, Princeton University Press,
Princeton, New Jersey.
Harvey, A. C. (1989). Forecasting, Structural Time Series Models and the
Kalman Filter, Cambridge University Press, Cambridge.
Harvey, A. C. (1993). Time Series Models, 2. edn, Harvester Wheatsheaf, New
York.
Hill, R. C., Knight, J. R. and Sirmans, C. F. (1997). Estimating Capital Asset
Price Indexes, Review of Economics and Statistics 79: 226“233.
13.6 Appendix 307
Koopman, S. J., Shepard, N. and Doornik, J. A. (1999). Statistical Algorithms
for Models in State Space Using SsfPack 2.2, Econometrics Journal 2: 107“
160.
Pe˜a, D., Tiao, G. C. and Tsay, R. S. (2001). A Course in Time Series Analysis,
n
Wiley, New York.
Schwann, G. M. (1998). A real estate price index for thin markets, Journal of
Real Estate Finance and Economics 16:3: 269“287.
Shiller, R. J. (1993). Macro Markets. Creating Institutions for Managing Soci
ety™s Largest Economic Risks, Clarendon Press, Oxford.
Shumway, R. H. and Sto¬er, D. S. (1982). An approach to time series smoothing
and forecasting using the EM algorithm, Journal of Time Series Analysis
3: 253“264.
Shumway, R. H. and Sto¬er, D. S. (2000). Time Series Analysis and Its Ap
plications, Springer, New York, Berlin.
Syds¦ter, K., Strøm, A. and Berck, P. (2000). Economists™ Mathematical Man
ual, 3. edn, Springer, New York, Berlin.
14 Long Memory E¬ects Trading
Strategy
Oliver Jim Blaskowitz and Peter Schmidt
14.1 Introduction
Long range dependence is widespread in nature and has been extensively doc
umented in economics and ¬nance, as well as in hydrology, meteorology, and
geophysics by authors such as Heyman, Tabatabai and Lakshman (1991), Hurst
(1951), Jones and Bri¬a (1992), Leland, Taqqu, Willinger and Wilson (1993)
and Peters (1994). It has a long history in economics and ¬nance, and has
remained a topic of active research in the study of ¬nancial time series, Beran
(1994).
Historical records of ¬nancial data typically exhibit distinct nonperiodical cycli
cal patterns that are indicative of the presence of signi¬cant power at low fre
quencies (i.e. long range dependencies). However, the statistical investigations
that have been performed to test for the presence of long range dependence in
economic time series representing returns of common stocks have often become
sources of major controversies. Asset returns exhibiting long range dependen
cies are inconsistent with the e¬cient market hypothesis, and cause havoc on
stochastic analysis techniques that have formed the basis of a broad part of
modern ¬nance theory and its applications, Lo (1991). In this chapter, we
examine the methods used in Hurst analysis, present a process exhibiting long
memory features, and give market evidence by applying Hurst™s R/S analysis
and ¬nally sketch a trading strategy for German voting and non“voting stocks.
310 14 Long Memory E¬ects Trading Strategy
14.2 Hurst and Rescaled Range Analysis
Hurst (1900“1978) was an English hydrologist, who worked in the early 20th
century on the Nile River Dam project. When designing a dam, the yearly
changes in water level are of particular concern in order to adapt the dam™s
storage capacity according to the natural environment. Studying an Egyptian
847“year record of the Nile River™s over¬‚ows, Hurst observed that ¬‚ood occur
rences could be characterized as persistent, i.e. heavier ¬‚oods were accompanied
by above average ¬‚ood occurrences, while below average occurrences were fol
lowed by minor ¬‚oods. In the process of this ¬ndings he developed the Rescaled
Range (R/S) Analysis.
We observe a stochastic process Yt at time points t ∈ I = {0, . . . , N }. Let n
be an integer that is small relative to N , and let A denote the integer part of
N/n. Divide the ˜interval™ I into A consecutive ˜subintervals™, each of length
n and with overlapping endpoints. In every subinterval correct the original
datum Yt for location, using the mean slope of the process in the subinterval,
obtaining Yt ’ (t/n) (Yan ’ Y(a’1)n ) for all t with (a ’ 1)n ¤ t ¤ an and for all
a = 1, . . . , A. Over the a™th subinterval Ia = {(a ’ 1)n, (a ’ 1)n + 1, . . . , an},
for 1 ¤ a ¤ A, construct the smallest box (with sides parallel to the coordinate
axes) such that the box contains all the ¬‚uctuations of Yt ’(t/n) (Yan ’Y(a’1)n )
that occur within Ia . Then, the height of the box equals
t
Yt ’ (Yan ’ Y(a’1)n )
Ra = max
n
(a’1)n¤t¤an
t
’ Yt ’ (Yan ’ Y(a’1)n )
min
n
(a’1)n¤t¤an
Figure 14.1 illustrates the procedure. Let Sa denote the empirical standard
error of the n variables Yt ’ Yt’1 , for (a ’ 1)n + 1 ¤ t ¤ an. If the process
Y is stationary then Sa varies little with a; in other cases, dividing Ra by
Sa corrects for the main e¬ects of scale inhomogeneity in both spatial and
temporal domains.
The total area of the boxes, corrected for scale, is proportional in n to
A
R Ra
’1
:= A . (2.1)
S Sa
n
a=1
ˆ
The slope H of the regression of log(R/S)n on log n, for k values of n, may be
taken as an estimator of the Hurst constant H describing longrange depen
dence of the process Y , Beran (1994) and Peters (1994).
14.2 Hurst and Rescaled Range Analysis 311
X(t)(t/n){X(an)X((a1)n)}
8
7.5
7
0 500 1000 1500 2000 2500
time t
Figure 14.1. The construction of the boxes in the R/S analysis.
If the process Y is stationary then correction for scale is not strictly necessary,
and we may take each Sa to be the constant 1. In that case the R“S statistic
ˆ
H is a version of the boxcounting estimator that is widely used in physical
science applications, Carter, Cawley and Mauldin (1988), Sullivan and Hunt
(1988) and Hunt (1990). The boxcounting estimator is related to the capacity
de¬nition of fractal dimension, Barnsley (1988) p. 172¬, and the R“S estimator
may be interpreted in the same way. Statistical properties of the boxcounting
estimator have been discussed by Hall and Wood (1993).
A more detailed analysis, exploiting dependence among the errors in the regres
sion of log(R/S)n on log n, may be undertaken in place of R“S analysis. See
Kent and Wood (1997) for a version of this approach in the case where scale
correction is unnecessary. However, as Kent and Wood show, the advantages
of the approach tend to be asymptotic in character, and sample sizes may need
to be extremely large before real improvements are obtained.
Hurst used the coe¬cient H as an index for the persistence of the time series
considered. For 0.5 < H < 1, it is positively persistent and characterized
by ˜long memory™ e¬ects, as described in the next section. A rather informal
interpretation of H used by practitioners is this: H may be interpreted as
the chance of movements with the same sign, Peters (1994). For H > 0.5,
it is more likely that an upward movement is followed by a movement of the
same (positive) sign, and a downward movement is more likely to be followed
312 14 Long Memory E¬ects Trading Strategy
by another downward movement. For H < 0.5, a downward movement is
more likely to be reversed by an upward movement thus implying the reverting
behavior.
14.3 Stationary Long Memory Processes
A stationary process X has the long memory property, if for its autocorrelation
function ρ(k) = Cov(Xi , Xi+k )/Var(X1 ) holds:
∞
= ∞.
ρ(k) (14.1)
k=’∞
That is, the autocorrelations decay to zero so slowly that their sum does not
converge, Beran (1994).
With respect to (14.1), note that the classical expression for the variance of the
¯ def n
sample mean, X = n’1 i=1 Xi , for independent and identically distributed
X1 , . . . , Xn ,
2
¯ = σ with σ 2 = Var(Xi )
Var(X) (14.2)
n
is not valid anymore. If correlations are neither zero and nor so small to be
¯
negligible, the variance of X is equal to
n’1
σ2 k
¯ 1’
Var(X) = 1+2 ρ(k) . (14.3)
n n
k=1
Thus, for long memory processes the variance of the sample mean converges to
zero at a slower rate than n’1 , Beran (1994). Note that long memory implies
positive long range correlations. It is essential to understand that long range
dependence is characterized by slowly decaying correlations, although nothing
is said about the size of a particular correlation at lag k. Due to the slow
decay it is sometimes di¬cult to detect non zero but very small correlations by
√
looking at the ±2/ n“con¬dence band. Beran (1994) gives an example where
√
the correct correlations are slowly decaying but within the ±2/ n“band. So
even if estimated correctly we would consider them as non signi¬cant.
Note that (14.1) holds in particular if the autocorrelation ρ(k) is approximately
ck’± with a constant c and a parameter ± ∈ (0, 1). If we know the autocor
14.3 Stationary Long Memory Processes 313
relations we also know the spectral density f (»), de¬ned as
∞
σ2
ρ(k)eik» .
f (») = (14.4)
2π
k=’∞
The structure of the autocorrelation then implies, that the spectral density is
approximately of the form cf k±’1 with a constant cf as » ’ 0. Thus the
spectral density has a pole at 0.
To connect the long memory property with the Hurst coe¬cient, we introduce
self similar processes. A stochastic process Yt is called self similar with self
similarity parameter H, if for any positive stretching factor c, the rescaled
process c’H Yct has the same distribution as the original process Yt . If the
increments Xt = Yt ’ Yt’1 are stationary, there autocorrelation function is
given by
1
k + 12H ’ 2k2H + k ’ 12H ,
ρ(k) =
2
Beran (1994). From a Taylor expansion of ρ it follows
ρ(k)
’ 1 for k ’ ∞ .
H(2H ’ 1)k 2H’2
This means, that for H > 0.5, the autocorrelation function ρ(k) is approxi
mately H(2H ’ 1)k ’± with ± = 2 ’ 2H ∈ (0, 1) and thus Xt has the long
memory property.
14.3.1 Fractional Brownian Motion and Noise
In this section, we introduce a particular self similar process with station
ary increments, namely the fractional Brownian motion (FBM) and fractional
Gaussian noise (FGN), Mandelbrot and van Ness (1968), Beran (1994).
DEFINITION 14.1 Let BH (t) be a stochastic process with continuous sam
ple paths and such that
• BH (t) is Gaussian
• BH (0) = 0
• E {BH (t) ’ BH (s)} = 0
314 14 Long Memory E¬ects Trading Strategy
σ2
t2H ’ t ’ s2H + s2H
• Cov {BH (t), BH (s)} = 2
for any H ∈ (0, 1) and σ 2 a variance scaling parameter. Then BH (t) is called
fractional Brownian motion.
Essentially, this de¬nition is the same as for standard Brownian motion besides
that the covariance structure is di¬erent. For H = 0.5, de¬nition 14.1 contains
standard Brownian motion as a special case but in general (H = 0.5), incre
ments BH (t) ’ BH (s) are not independent anymore. The stochastic process
resulting by computing ¬rst di¬erences of FBM is called FGN with parameter
H. The covariance at lag k of FGN follows from de¬nition 14.1:
Cov {BH (t) ’ BH (t ’ 1), BH (t + k) ’ BH (t + k ’ 1)}
γ(k) =
σ2
k + 12H ’ 2k2H + k ’ 12H
= (14.5)
2
For 0.5 < H < 1 the process has long range dependence, and for 0 < H < 0.5
the process has short range dependence.
Figures 14.2 and 14.3 show two simulated paths of N = 1000 observations of
FGN with parameter H = 0.8 and H = 0.2 using an algorithm proposed by
Davies and Harte (1987). For H = 0.2, the FBM path is much more jagged
and the range of the y“axis is about ten times smaller than for H = 0.8 which
is due to the reverting behavior of the time series.
The estimated autocorrelation function (ACF) for the path simulated with
√
H = 0.8 along with the ±2/ N “con¬dence band is shown in Figure 14.4.
For comparison the ACF used to simulate the process given by (14.5) is su
perimposed (dashed line). The slow decay of correlations can be seen clearly.
Applying R/S analysis we can retrieve the Hurst coe¬cient used to simulate
the process. Figure 14.5 displays the estimated regression line and the data
points used in the regression. We simulate the process with H = 0.8 and the
ˆ
R/S statistic yields H = 0.83.
Finally, we mention that fractional Brownian motion is not the only stationary
process revealing properties of systems with long memory. Fractional ARIMA
processes are an alternative to FBM, Beran (1994). As well, there are non
stationary processes with in¬nite second moments that can be used to model
long range dependence, Samrodnitsky and Taqqu (1994).
14.4 Data Analysis 315
Sim. FGN with N=1000, H=0.80
2 0 2
0 250 500 750 1000
t
Sim. FBM with N=1000, H=0.80
200
0
0 250 500 750 1000
t
Figure 14.2. Simulated FGN with H = 0.8, N = 1000 and path of
corresponding FBM.
Sim. FGN with N=1000, H=0.20
4
0
4
0 250 500 750 1000
t
Sim. FBM with N=1000, H=0.20
5 0 5
0 250 500 750 1000
t
Figure 14.3. Simulated FGN with H = 0.2, N = 1000 and path of
corresponding FBM. XFGSimFBM.xpl
14.4 Data Analysis
A set of four pairs of voting and non“voting German stocks will be subject to
our empirical analysis. More precisely, our data sample retrieved from the data
information service Thompson Financial Datastream, consists of 7290 daily
316 14 Long Memory E¬ects Trading Strategy
True (dashed) & Est. ACF of Sim. FGN: N=1000, H=0.80
1
0.5
ACF
0
0 20 40 60 80
lag k
Figure 14.4. Estimated and true ACF of FGN simulated with H = 0.8,
N = 1000. XFGSimFBM.xpl
HurstPlot: FGN(N=1000, H=0.80), est. H=0.83
2 2.5 3 3.5 4 4.5
log[E(R/S)]
3 3.5 4 4.5 5 5.5 6
log(k)
Figure 14.5. Hurst regression and estimated Hurst coe¬cient
ˆ
(H = 0.83) of FBM simulated with H = 0.8, N = 1000.
XFGSimFBMHurst.xpl
closing prices of stocks of WMF, Dyckerho¬, KSB and RWE from January 01,
1973, to December 12, 2000.
Figure 14.6 shows the performance of WMF stocks in our data period. The
plot indicates an intimate relationship of both assets. Since the performance
of both kinds of stocks are in¬‚uenced by the same economic underlyings, their
relative value should be stable over time. If this holds, the log“di¬erence Xt of
v nv
the pairs of voting (St ) and non“voting stocks (St ),
def v nv
Xt = log St ’ log St (14.6)
14.4 Data Analysis 317
should exhibit a reverting behavior and therefore an R/S analysis should yield
estimates of the Hurst coe¬cient smaller than 0.5. In order to reduce the num
ber of plots we show only the plot of WMF stocks. One may start the quantlet
XFGStocksPlots.xpl to see the time series for the other companies as well.
First, we perform R/S analysis on both individual stocks and the voting/non“
voting log“di¬erences. In a second step, a trading strategy is applied to all four
voting/non“voting log“di¬erences.
Time Series of Voting(dashed) and Non Voting WMF Stocks
EUR
35
25
15
5
1973 1980 1990 2000
Time
Figure 14.6. Time series of voting and non“voting WMF stocks.
XFGStocksPlots.xpl
Table 14.1 gives the R/S statistic of each individual stock and of the log“
ˆ
di¬erence process of voting and non“voting stocks. While H is close to 0.5
for each time series taken separately, we ¬nd for the log di¬erences a Hurst
coe¬cient indicating negative persistence, i.e. H < 0.5.
WMF Dyck. KSB RWE
nv v nv v nv v nv v
Stock 0.51 0.53 0.57 0.52 0.53 0.51 0.50 0.51
Di¬erences 0.33 0.37 0.33 0.41
Table 14.1. Estimated Hurst coe¬cients of each stock and of log“
di¬erences.
318 14 Long Memory E¬ects Trading Strategy
To test for the signi¬cance of the estimated Hurst coe¬cients we need to know
the ¬nite sample distribution of the R/S statistic. Usually, if the probabilistic
behavior of a test statistic is unknown, it is approximated by its asymptotic
distribution when the number of observations is large. Unfortunately, as, for
example, Lo (1991) shows, such an asymptotic approximation is inaccurate in
the case of the R/S statistic. This problem may be solved by means of bootstrap
and simulation methods. A semiparametric bootstrap approach to hypothesis
testing for the Hurst coe¬cient has been introduced by Hall, H¨rdle, Kleinow
a
and Schmidt (2000), In the spirit of this chapter we use Brownian motion
(H = 0.5) to simulate under the null hypothesis. Under the null hypothesis
the log“di¬erence process follows a standard Brownian motion and by Monte
Carlo simulation we compute 99%, 95% and 90% con¬dence intervals of the
R/S statistic. The results are given in Table 14.2. While the estimated Hurst
coe¬cients for each individual stock are at least contained in the 99% con¬dence
interval, we consider the R/S statistic for voting/non“voting log di¬erences as
signi¬cant.
N Mean 90% 95% 99%
7289 0.543 [0.510, 0.576] [0.504, 0.582] [0.491, 0.595]
Table 14.2. Simulated con¬dence intervals for R/S statistic for Brown
ian motion.
14.5 Trading the Negative Persistence
The data analysis conducted so far indicates a negative persistence (H < 0.5)
of the log di¬erences of pairs of voting and non“voting stocks of a company. It
should be possible to take advantage of this knowledge. If we found a pro¬table
trading strategy, we would interpret this result as a further indication for the
reverting behavior of voting/non“voting log“di¬erences.
The average relationship between voting and non“voting stocks in the sample
period may be expressed in the following way,
log(voting) = β log(nonvoting) + µ
where β may be estimated by linear regression. If the log“di¬erences of voting
and non“voting stocks are reverting as the R/S analysis indicates, negative
di¬erences, Xt < 0, are often followed by positive di¬erences and vice versa.
In terms of the Hurst coe¬cient interpretation, given a negative di¬erence, a
14.5 Trading the Negative Persistence 319
positive di¬erence has a higher chance to appear in the future than a negative
one and vice versa, implying voting stocks probably to become relatively more
expensive than their non“voting counterparts. Thus, we go long the voting and
short the non“voting stock. In case of the inverse situation, we carry out the
inverse trade (short voting and long non“voting). When initiating a trade we
take a cash neutral position. That is, we go long one share of the voting and
sell short m shares of the non“voting stock to obtain a zero cash ¬‚ow from this
action.
But how to know that a ˜turning point™ is reached? What is a signal for the
reverse? Naturally, one could think, the longer a negative di¬erence persisted,
the more likely the di¬erence is going to be positive. In our simulation, we cal
culate the maximum and minimum di¬erence of the preceding M trading days
(for example M = 50, 100, 150). If the current di¬erence is more negative than
the minimum over the last M trading days, we proceed from the assumption
that a reverse is to come and that the di¬erence is going to be positive, thereby
triggering a long voting and short non“voting position. A di¬erence greater
than the M day maximum releases the opposite position.
When we take a new position, we compute the cash ¬‚ow from closing the old
one. Finally, we calculate the total cash ¬‚ow, i.e. we sum up all cash ¬‚ows
without taking interests into account. To account for transaction costs, we
compute the total net cash ¬‚ow. For each share bought or sold, we calculate a
hypothetical percentage, say 0.5%, of the share price and subtract the sum of
all costs incurred from the total cash ¬‚ow. In order to compare the total net
cash ¬‚ows of our four pairs of stocks which have di¬erent levels of stock prices,
we normalize them by taking WMF stocks as a numeraire.
In Table 14.3 we show the total net cash ¬‚ows and in Table 14.4 the number
of trade reverses are given. It is clear that for increasing transaction costs the
performance deteriorates, a feature common for all 4 pairs of stocks. Moreover,
it is quite obvious that the number of trade reverses decreases with the number
of days used to compute the signal. An interesting point to note is that for
RWE, which is in the German DAX30, the total net cash ¬‚ow is worse in all
situations. A possible explanation would be that since the Hurst coe¬cient
is the highest, the log“di¬erences contain less ˜reversion™. Thus, the strategy
designed to exploit the reverting behavior should perform rather poorly. WMF
and KSB have a smaller Hurst coe¬cient than RWE and the strategy performs
320 14 Long Memory E¬ects Trading Strategy
better than for RWE. Furthermore, the payo¬ pattern is very similar in all
situations. Dyckerho¬ with a Hurst coe¬cient of H = 0.37 exhibits a payo¬
structure that rather resembles the one of WMF/KSB.
Transaction M WMF Dyckerho¬ KSB RWE
Costs H = 0.33 H = 0.37 H = 0.33 H = 0.41
0.00 50 133.16 197.54 138.68 39.93
100 104.44 122.91 118.85 20.67
150 71.09 62.73 56.78 8.80
0.005 50 116.92 176.49 122.32 21.50
100 94.87 111.82 109.26 12.16
150 64.78 57.25 51.86 2.90
0.01 50 100.69 155.43 105.96 3.07
100 85.30 100.73 99.68 3.65
’3.01
150 58.48 51.77 49.97
Table 14.3. Performance of Long Memory Strategies (TotalNetCash
Flow in EUR). XFGLongMemTrade.xpl
M WMF Dyckerho¬ KSB RWE
50 120 141 132 145
100 68 69 69 59
150 47 35 41 42
Table 14.4. Number of Reverses of Long Memory Trades
Regarding the interpretation of the trading strategy, one has to be aware that
neither the cash ¬‚ows are adjusted for risk nor did we account for interest rate
e¬ects although the analysis spread over a period of time of about 26 years.
Bibliography
Barnsley, M. (1988). Fractals everywhere., Boston, MA etc.: Academic Press,
Inc.
Beran, J. (1994). Statistics for Long Memory Processes, Chapman and Hall,
New York.
14.5 Trading the Negative Persistence 321
Carter, P., Cawley, R. and Mauldin, R. (1988). Mathematics of dimension mea
surements of graphs of functions, in D. Weitz, L. Sander and B. Mandel
brot (eds), Proc. Symb. Fractal Aspects of Materials, Disordered Systems,
pp. 183“186.
Davies, R. B. and Harte, D. S. (1987). Test for Hurst E¬ect, Biometrica 74: 95“
102.
Hall, P., H¨rdle, W., Kleinow, T. and Schmidt, P. (2000). Semiparametric
a
bootstrap approach to hypothesis tests and con¬dence intervals for the
hurst coe¬cient, Statistical Inference for stochastic Processes 3.
Hall, P. and Wood, A. (1993). On the performance of boxcounting estimators
of fractal dimension., Biometrika 80(1): 246“252.
Heyman, D., Tabatabai, A. and Lakshman, T.V. (1993). Statistical analysis
and simulation of video teleconferencing in ATM networks, IEEE Trans.
Circuits. Syst. Video Technol., 2, 49“59.
Hunt, F. (1990). Error analysis and convergence of capacity dimension algo
rithms., SIAM J. Appl. Math. 50(1): 307“321.
Hurst, H. E. (1951). Long Term Storage Capacity of Reservoirs, Trans. Am.
Soc. Civil Engineers 116, 770“799.
Jones, P.D. and Bri¬a, K.R. (1992). Global surface air temperature variations
during the twentieth century: Part 1, spatial, temporal and seasonals
details, The Holocene 2, 165“179.
Kent, J. T. and Wood, A. T. (1997). Estimating the fractal dimension of a
locally selfsimilar Gaussian process by using increments., J. R. Stat. Soc.,
Ser. B 59(3): 679“699.
Leland, W.E., Taqqu, M.S., Willinger, W. and Wilson, D.V. (1993). Ethernet
tra¬c is self“similar: Stochastic modelling of packet tra¬c data, preprint,
Bellcore, Morristown.
Lo, A.W. (1991). Longterm memory in stock market prices, Econometrica,
59, 1279“1313.
Mandelbrot, B.B. and van Ness, J.W. (1968). Fractional Brownian Motion,
fractional Noises and Applications, SIAM Rev.10, 4, 422“437.
Peters, E.E. (1994). Fractal Market Analysis: Applying Chaos Theory to In
vestment and Economics, John Wiley & Sons, New York.
322 14 Long Memory E¬ects Trading Strategy
Samrodnitsky, G. and Taqqu, M.S. (1994). Stable nonGaussian Random Pro
cesses: Stochastic Models with in¬nite variance, Chapman and Hall, New
York.
Sullivan, F. and Hunt, F. (1988). How to estimate capacity dimension, Nuclear
Physics B (Proc. Suppl.) pp. 125“128.
15 Locally time homogeneous time
series modeling
Danilo Mercurio
15.1 Intervals of homogeneity
An adaptive estimation algorithm for time series is presented in this chapter.
The basic idea is the following: given a time series and a linear model, we
select online the largest sample of the most recent observations, such that
the model is not rejected. Assume for example that the data can be well
¬tted by a regression, an autoregression or even by a constant in an unknown
interval. The main problem is then to detect the time interval where the model
approximately holds. We call such an interval: interval of time homogeneity.
This approach appears to be suitable in ¬nancial econometrics, where an on
line analysis of large data sets, like e.g. in backtesting, has to be performed. In
this case, as soon as a new observation becomes available, the model is checked,
the sample size is optimally adapted and a revised forecast is produced.
In the remainder of the chapter we brie¬‚y present the theoretical foundations
of the proposed algorithm which are due to Liptser and Spokoiny (1999) and
we describe its implementation. Then, we provide two applications to ¬nancial
data. In the ¬rst one we estimate the possibly time varying coe¬cients of an
exchange rate basket, while in the second one the volatility of an exchange rate
time series is ¬tted to a locally constant model. The main references can be
found in H¨rdle, Herwartz and Spokoiny (2001), Mercurio and Spokoiny (2000),
a
H¨rdle, Spokoiny and Teyssi`re (2000) and Mercurio and Torricelli (2001).
a e
Let us consider the following linear regression equation:
Yt = Xt θ + σµt , t = 1, . . . , T (15.1)
324 15 Locally time homogeneous time series modeling
where Yt is real valued, Xt = (X1,t . . . Xp,t ) and θ = (θ1 . . . θp ) are Rp
valued and µt is a standard normally distributed random variable. If the matrix
T
t=1 Xt Xt is nonsingular with inverse W , then the least squares estimator of
θ is:
T
θ=W Xt Yt . (15.2)
t=1
De¬ne wkk as the kth element on the diagonal of W and let » be a positive
scalar. For nonrandom regressors,the following exponential probability bound
is easy to prove:
√ »2
P(θk ’ θk  > »σ wkk ) ¤ 2e’ 2 , k = 1, . . . , p. (15.3)
Indeed, the estimation error θk ’ θk is N(0, wkk σ 2 ) distributed, therefore:
2
»(θk ’ θk ) »2
’
√
1 = E exp
σ wkk 2
»(θk ’ θk ) »2 √
≥ E exp ’ 1(θk ’ θk > »σ wkk )
√
σ wkk 2
»2 √
≥ exp P(θk ’ θk > »σ wkk ).
2
The result in (15.3) follows from the symmetry of the normal distribution.
Equation (15.3) has been generalized by Liptser and Spokoiny (1999) to the
case of nonrandom regressors. More precisely, they allow the Xt to be only con
ditionally independent of µt , and they include lagged values of Yt as regressors.
In this case the bound reads roughly as follows:
√ 2
’»
P(θk ’ θk  > »σ wkk ; W is nonsingular ) ¤ P(»)e . (15.4)
2
Where P(») is a polynomial in ». It must be noticed that (15.4) is not as sharp
as (15.3), furthermore, because of the randomness of W , (15.4) holds only on
the set where W is nonsingular, nevertheless this set has in many cases a large
probability. For example when Yt follows an ergodic autoregressive process and
the number of observations is at least moderately large. More technical details
are given in Section 15.4.
We now describe how the bound (15.4) can be used in order to estimate the
coe¬cients θ in the regression equation (15.1) when the regressors are (possi
bly) stochastic and the coe¬cients are not constant, but follow a jump process.
15.1 Intervals of homogeneity 325
θi,t 6

„ ’m „
time
Figure 15.1. Example of a locally homogeneous process.
The procedure that we describe does not require an explicit expression of the
law of the process θt , but it only assumes that θt is constant on some unknown
time interval I = [„ ’ m, „ ], „ ’ m > 0, „, m ∈ N. This interval is referred
as an interval of time homogeneity and a model which is constant only on some
time interval is called locally time homogeneous.
Let us now de¬ne some notation. The expression θ„ will describe the (¬ltering)
estimator of the process (θt )t∈N at time „ ; that is to say, the estimator which
uses only observations up to time „ . For example if θ is constant, the recursive
estimator of the form:
’1
„ „
θ„ = Xs Xs Xs Ys ,
s=1 s=1
represents the best linear estimator for θ. But, if the coe¬cients are not con
stant and follow a jump process, like in the picture above a recursive estimator
cannot provide good results. Ideally, only the observations in the interval
I = [„ ’ m, „ ] should be used for the estimation of θ„ . Actually, an estima
tor of θ„ using the observation of a subinterval J ‚ I would be less e¬cient,
while an estimator using the observation of a larger interval K ⊃ I would be
biased. The main objective is therefore to estimate the largest interval of time
homogeneity. We refer to this estimator as I = [„ ’ m, „ ]. On this interval I
326 15 Locally time homogeneous time series modeling
we estimate θ„ with ordinary least squares (OLS):
’1
«
θ„ = θI = Xs Xs Xs Ys . (15.5)
s∈I s∈I
In order to determine I we use the idea of pointwise adaptive estimation de
scribed in Lepski (1990), Lepski and Spokoiny (1997) and Spokoiny (1998).
The idea of the method can be explained as follows.
Suppose that I is an intervalcandidate, that is, we expect timehomogeneity in
I and hence in every subinterval J ‚ I. This implies that the mean values of the
θI and θJ nearly coincide. Furthermore, we know on the basis of equation (15.4)
that the events
√ √
θi,I ’ θ„  ¤ µσ wii,I and θi,J ’ θ„  ¤ »σ wii,J
occur with high probability for some su¬ciently large constants » and µ. The
adaptive estimation procedure therefore roughly corresponds to a family of
tests to check whether θI does not di¬er signi¬cantly from θJ . The latter is
done on the basis of the triangle inequality and of equation (15.4) which assigns
a large probability to the event
√ √
θi,I ’ θi,J  ¤ µσ wii,I + »σ wii,J
under the assumption of homogeneity within I, provided that µ and » are
su¬ciently large. Therefore, if there exists an interval J ‚ I such that the
hypothesis θi,I = θi,J cannot be accepted, we reject the hypothesis of time
homogeneity for the interval I. Finally, our adaptive estimator corresponds to
the largest interval I such that the hypothesis of homogeneity is not rejected
for I itself and all smaller intervals.
15.1.1 The adaptive estimator
Now we present a formal description. Suppose that a family I of interval
candidates I is ¬xed. Each of them is of the form I = [„ ’ m, „ ], so that the
set I is ordered due to m. With every such interval we associate an estimate
θi,I of the parameter θi,„ and the corresponding conditional standard deviation
√
wii,I . Next, for every interval I from I, we suppose to be given a set J (I)
of testing subintervals J. For every J ∈ J (I), we construct the corresponding
√
estimate θi,J from the observations for t ∈ J and compute wii,J . Now, with
15.1 Intervals of homogeneity 327
two constants µ and », de¬ne the adaptive choice of the interval of homogeneity
by the following iterative procedure:
• Initialization: Select the smallest interval in I
• Iteration: Select the next interval I in I and calculate the corresponding
√
estimate θi,I and the conditional standard deviation wii,I σ
• Testing homogeneity: Reject I, if there exists one J ∈ J (I), and i =
1, . . . , p such that
√ √
θi,I ’ θi,J  > µσ wii,I + »σ wii,J . (15.6)
• Loop: If I is not rejected, then continue with the iteration step by choos
ing a larger interval. Otherwise, set I = “the latest non rejected I”.
The adaptive estimator θ„ of θ„ is de¬ned by applying the selected interval I:
θi,„ = θi,I for i = 1, . . . , p.
As for the variance estimation, note that the previously described procedure re
quires the knowledge of the variance σ 2 of the errors. In practical applications,
σ 2 is typically unknown and has to be estimated from the data. The regression
representation (15.1) and local time homogeneity suggests to apply a residual
based estimator. Given an interval I = [„ ’ m, „ ], we construct the parameter
estimate θI . Next the pseudoresiduals µt are de¬ned as µt = Yt ’Xt θI . Finally
the variance estimator is de¬ned by averaging the squared pseudoresiduals:
1
σ2 = µ2 .
t
I
t∈I
15.1.2 A small simulation study
The performance of the adaptive estimator is evaluated with data from the
following process:
Yt = θ1,t + θ2,t X2,t + θ3,t X3,t + σµt .
The length of the sample is 300. The regressors X2 and X3 are two independent
random walks. The regressor coe¬cients are constant in the ¬rst half of the
328 15 Locally time homogeneous time series modeling
1 ¤ t ¤ 150 151 ¤ t ¤ 300
large jump medium jump small jump
θ1,t = 1 θ1,t = .85 θ1,t = .99 θ1,t = .9995
θ2,t = .006 θ2,t = .0015 θ2,t = .004 θ2,t = .0055
θ3,t = .025 θ3,t = .04 θ3,t = .028 θ3,t = .0255
Table 15.1. Simulated models.
sample, then they make a jump after which they continue being constant until
the end of the sample. We simulate three models with jumps of di¬erent
magnitude. The values of the simulated models are presented in Table 15.1.
The error term µt is a standard Gaussian white noise, and σ = 10’2 . Note that
the average value of σµt  equals 10’2 2/π ≈ 0.008, therefore the small jump
of magnitude 0.0005 is clearly not visible by eye. For each of the three models
above 100 realizations of the white noise µt are generated and the adaptive
estimation is performed.
In order to implement the procedure we need two parameters: µ and », and two
sets of intervals: I and J (I). As far as the latter are concerned the simplest
proposal is to use a regular grid G = {tk } with tk = m0 k for some integer
m0 and with „ = tk— belonging to the grid. We next consider the intervals
Ik = [tk , tk— [= [tk , „ [ for all tk < tk— = „ . Every interval Ik contains exactly
k — ’ k smaller intervals J = [tk , tk— [. So that for every interval Ik = [tk , tk— [
and k : k < k < k — we de¬ne the set J (Ik ) of testing subintervals J by
taking all smaller intervals with right end point tk— : J = [tk , tk— [ and all
smaller intervals with left end point tk :J = [tk , tk [:
J (Ik ) = {J = [tk , tk— [ or J = [tk , tk [: k < k < k — }.
The testing interval sets I and J (I) are therefore identi¬ed by the parameter
m0 : the grid step.
We are now left with the choice of three parameters: », µ and m0 . These
parameters act as the smoothing parameters in the classical nonparametric
estimation. The value of m0 determines the number of points at which the
time homogeneity is tested and it de¬nes the minimal delay after which a jump
15.2 Estimating the coe¬cients of an exchange rate basket 329
can be discovered. Simulation results have shown that small changes of m0
do not essentially a¬ect the results of the estimation and, depending on the
number of parameters to be estimated, it can be set between 10 and 50.
The choice of » and µ is more critical because these parameters determine the
acceptance or the rejection of the interval of time homogeneity as it can be
seen from equation (15.6). Large values of » and µ reduce the sensitivity of
the algorithm and may delay the detection of the change point, while small
values make the procedure more sensitive to small changes in the values of the
estimated parameters and may increase the probability of a typeI error.
For the simulation, we set: m0 = 30, » = 2 and µ = 4, while a rule for the
selection of » and µ for real application will be discussed in the next section.
Figure 15.2 shows the results of the simulation. The true value of the coe¬cients
is plotted (θ1,t : ¬rst row, θ2,t : second row, θ3,t : third row) along with the
median, the maximum and the minimum of the estimates from all realizations
for each model at each time point. The simulation results are very satisfactory.
The change point is quickly detected, almost within the minimal delay of 30
periods for all three models, so that the adaptive estimation procedure show a
good performance even for the small jump model.
15.2 Estimating the coe¬cients of an exchange
rate basket
In this section we compare the adaptive estimator with standard procedures
which have been designed to cope with time varying regressor coe¬cients. A
simple solution to this problem consists in applying a window estimator, i.e.
an estimator which only uses the most recent k observations:
’1
t t
θt = Xs Xs Xs Ys , (15.7)
s=t’k s=t’k
where the value of k is speci¬ed by the practitioner. Another, more re¬ned
technique, consists in describing the coe¬cients θ as an unobserved stochastic
process: (θt )t∈N , see Elliot, Aggoun and Moore (1995). Apart from the cases
when there is some knowledge about the data generating process of θt , the
most common speci¬cation is as a multivariate random walk:
ζt ∼ N(0, Σ).
θt = θt’1 + ζt (15.8)
330 15 Locally time homogeneous time series modeling
LARGE JUMP MEDIUM JUMP SMALL JUMP
50 100 150 200 250 300 50 100 150 200 250 300 50 100 150 200 250 300
ALPHA_1 ALPHA_1 ALPHA_1
50 100 150 200 250 300 50 100 150 200 250 300 50 100 150 200 250 300
ALPHA_2 ALPHA_2 ALPHA_2
50 100 150 200 250 300 50 100 150 200 250 300 50 100 150 200 250 300
ALPHA_3 ALPHA_3 ALPHA_3
Figure 15.2. Online estimates of the regression coe¬cients with jumps
of di¬erent magnitude. Median (thick dotted line), maximum and min
imum (thin dotted line) among all estimates.
In this context, equations (15.8) and (15.1) can be regarded as a state space
model, where equation (15.8) is the state equation (the signal) and equation
(15.1) is the measurement equation and it plays the role of a noisy observation
of θt . A Kalman ¬lter algorithm can be used for the estimation, see Cooley
and Prescott (1973). The Kalman ¬lter algorithm requires the initialization of
15.2 Estimating the coe¬cients of an exchange rate basket 331
two variables: θ00 and P00 = Cov(θ00 ) and its recursions read as follows, see
Chui and Chen (1998):
±
P00 = Cov(θ00 )
tt’1 = Pt’1t’1 + Σσ 2
P
= Ptt’1 Xt (Xt Ptt’1 Xt + σ 2 )’1
G
t
= (I ’ Gt Xt )Ptt’1
Ptt
tt’1 = θt’1t’1
θ
= θtt’1 + Gt (Yt ’ Xt θtt’1 ).
θtt
The question of the initialization of the Kalman ¬lter will be discussed in
the next section together with the Thai Baht basket example. In the notation
above, the index tt’1 denotes the estimate performed using all the observation
before time t (forecasting estimate), while tt refers to the estimate performed
using all the observations up to time t (¬ltering estimate). The four estimators
described above: the adaptive, the recursive, the window and the Kalman ¬lter
Estimator are now applied to the data set of the Thai Baht basket. For deeper
analysis of these data see Christo¬ersen and Giorgianni (2000) and Mercurio
and Torricelli (2001).
15.2.1 The Thai Baht basket
An exchange rate basket is a form of pegged exchange rate regime and it takes
place whenever the domestic currency can be expressed as a linear combination
of foreign currencies. A currency basket can be therefore expressed in the form
of equation (15.1), where: X1,t is set constantly equal to one and is taken as
numeraire, Yt represents the home currency exchange rate with respect to the
numeraire, and Xj,t is the amount of currency 1 per unit of currency j, i.e.
the cross currency exchange rate. The above relationship usually holds only on
the average, because the central bank cannot control the exchange rate exactly,
therefore the error term µt is added.
Because modern capital mobility enables the investors to exploit the interest
rate di¬erentials which may arise between the domestic and the foreign cur
rencies, a pegged exchange rate regime can become an incentive to speculation
and eventually lead to destabilization of the exchange rate, in spite of the fact
that its purpose is to reduce exchange rate ¬‚uctuations, see Eichengreen, Mas
son, Savastano and Sharma (1999). Indeed, it appears that one of the causes
which have led to the Asian crisis of 1997 can be searched in short term capital
investments.
332 15 Locally time homogeneous time series modeling
From 1985 until its suspension on July 2, 1997 (following a speculative attack)
the Bath was pegged to a basket of currencies consisting of Thailand™s main
trading partners. In order to gain greater discretion in setting monetary pol
icy, the Bank of Thailand neither disclosed the currencies in the basket nor the
weights. Uno¬cially, it was known that the currencies composing the basket
were: US Dollar, Japanese Yen and German Mark. The fact that the public
was not aware of the values of the basket weights, also enabled the monetary
authorities to secretly adjust their values in order to react to changes in eco
nomic fundamentals and/or speculative pressures. Therefore one could express
the USD/THB exchange rate in the following way:
YU SD/T HB,t = θU SD,t + θDEM,t XU SD/DEM,t + θJP Y,t XU SD/JP Y,t + σµt .
This exchange rate policy had provided Thailand with a good stability of the
exchange rate as it can be seen in Figure 15.3. During the same period, though,
the interest rates had maintained constantly higher than the ones of the coun
tries composing the basket, as it is shown in Figure 15.4.
This facts suggest the implementation of a speculative strategy, which con
sists in borrowing from the countries with a lower interest rate and lending
to the ones with an higher interest rate. A formal description of the problem
can be made relying on a meanvariance hedging approach, see Musiela and
— —
Rutkowski (1997). The optimal investment strategy ξ1 , . . . , ξp is obtained by
the minimization of the quadratic cost function below:
±«
2
p
E Yt+h ’ ξj Xj,t+h Ft .
j=1
The solution is:
—
ξj = E(θj,t+h Ft ) for j = 1, . . . , p.
It can be seen that, when the interest rates in Thailand (r0 ) are su¬ciently
high with respect to the foreign interest rates (rj , j = 1, . . . , p) the following
inequality holds
p
’1
(1 + rj )’1 E(θj,t+h Ft )Xj,t .
(1 + r0 ) Yt < (15.9)
j=1
This means that an investment in Thailand is cheaper than an investment
with the same expected revenue in the countries composing the basket. In the
15.2 Estimating the coe¬cients of an exchange rate basket 333
DEM/USD
1.7
1.6
Y
1.5
1.4
1992 1993 1994 1995 1996 1997
X
JPY/USD
130
120
110
Y
100
90
80
1992 1993 1994 1995 1996 1997
X
THB/USD
26
25.5
Y
25
24.5
1992 1993 1994 1995 1996 1997
X
Figure 15.3. Exchange rate time series. XFGbasket.xpl
empirical analysis we ¬nd out that the relationship (15.9) is ful¬lled during
the whole period under investigation for any of the four methods that we use
to estimate the basket weights. Therefore it is possible to construct a mean
334 15 Locally time homogeneous time series modeling
1 month interest rates
10
Y*E2