Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Journal of Financial Econometrics, 2005, Vol. 3, No.

3, 422–441

The Stability of Factor Models of


Interest Rates
Francesco Audrino
University of Lugano
Giovanni Barone-Adesi

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


University of Lugano
Antonietta Mira
University of Insubria

abstract
The daily term structure of interest rates is filtered to reduce the influence of
cross-correlations and autocorrelations on its factors. A three-factor model is
fitted to the filtered data. We perform statistical tests, finding that factor loadings
are unstable through time for daily data. This finding is not due to the presence
of outliers nor to the selected number of factors. Such an instability problem can
be solved when applying the factor analysis on multivariate scaled residuals,
filtered using a nonparametric technique based on functional gradient descent.

keywords: factor analysis, FGD, robust regression, term structure

This article analyzes the stability of the daily factor structure of interest rates.
Although daily interest rates exhibit an unstable structure, suitably filtered inno-
vations appear to be well described by three factors with stable loadings. Since
our factor structure is constructed on filtered innovations (independent over time)
and not on raw interest rate data, stable factors cannot be interpreted as level,
slope, and curvature as in most studies. Our use of daily data departs from the
more common use of monthly data. The motivation for our study is that the
stability of factor loadings is necessary for effective management of interest rate
risk. In fact, daily interest rate changes can then be described as functions of
underlying factor scores, which need to be unpredictable to be interpreted as
sources of risk.

Financial support from the National Centre of Competence in Research ‘‘Financial Valuation and Risk
Management’’ (NCCR FINRISK) and from F.A.R. 2004 of the Universtiy of Insubria is gratefully acknowl-
edged. Address correspondence to Autonietta Mira, Department of Economics, University of Insubria, Via
Ravasi 2, I-21100 Varese, Italy, or e-mail: antonietta.mira@uninsubriait.

doi:10.1093/jjfinec/nbi019
ª The Author 2005. Published by Oxford University Press. All rights reserved. For permissions,
please e-mail: journals.permissions@oupjournals.org.
Audrino et al. | The Stability of Factor Models 423

Factor models of the term structure proposed to date assume the stability of
underlying factors and their loadings without testing them. Our study is the first
study to apply a high-dimensional filter, based on functional gradient descent
[see Friedman, Hastie, and Tibshirani (2000) and Friedman (2001)], to remove
autocorrelations and cross-correlations from interest rate innovations. We per-
form formal tests of stability on filtered innovations, finding that three factors are
sufficient to describe them. Our work complements the large number of studies
that in the last decade have provided some insight into the number and the nature
of common factors needed to describe the dynamic evolution of the term structure

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


of interest rates.1
Recently some researchers tried to detect whether the factor decomposition
obtained from a factor analysis of the term structure of interest rates was stable
through time. To this purpose, Bliss (1997) divides his sample into three subper-
iods. He notices that, although factor volatility varies across subperiods, factor
loadings show a consistent pattern across different subperiods.2 As it appears
empirically that factor loadings remain fairly constant across subperiods, but the
volatility explained by the factors fluctuates over time, more recent studies
propose the use of common principal components analysis to estimate the factors
driving the term structure of interest rates in the presence of a time-varying
covariance structure. In this context, Perignon and Villa (2002) associate, for the
first time, groups to successive time periods. In their empirical investigations of
monthly bond yield changes, they found that the variance accounted for by the
first three factors changes substantially from subperiod to subperiod. However,
since factor loadings are fairly constant across subperiods, the factor structure
does not change appreciably.3
The conclusion that a factor decomposition or a (common) principal compo-
nent decomposition is robust through time, given in the above mentioned studies,
is not supported by any formal statistical test. More recently, Perignon and Villa
(2004) provide some key insight into the stability of the latent factor structure of
interest rates through time, based on a formal testing procedure. Similarly, in this
article we perform statistical tests investigating the time robustness of factor
decompositions of the term structure of interest rates. To this aim we perform a
robust three-factor analysis of the independent filtered innovations, allowing for
nongaussianity of the innovations [see Croux et al., (2003)]. Then we divide the
period under consideration into four different subperiods following Perignon and
Villa (2002). For each maturity, we regress the innovation series using as pre-
dictors the vectors of factor scores multiplied by the corresponding factor

1
See for example Stambaugh (1988), Steeley (1990), Carverhill and Strickland (1992), and Litterman and
Scheinkman (1991) for a factor analysis of the U.S. term structure. A critical reexamination of classical
results of factor analysis of interest rates has been proposed by Leikkos (2000).
2
Phoa (2000) and Chapman and Pearson (2001) also found, in their empirical investigations, that the factor
decomposition of the U.S. term structure seems to be robust through time.
3
Analogous results were found by Diebold and Li (2002) in a different setting, that is, the Nelson-Siegel
framework, for the monthly forecasts of the U.S. term structure of government zero-coupon bond yield
levels and changes.
424 Journal of Financial Econometrics

loadings and some dummy variables. We perform tests based on the hypothesis
that the regression coefficients (loadings) in the different subperiods are equal (or,
in other words, that coefficients of the dummy variables are equal to zero). Using
this strategy, we are able to detect changes in factor loadings.
In contrast with the common belief and the empirical results found by Bliss
(1997), Perignon and Villa (2002), and Diebold and Li (2002) for monthly data, our
statistical investigation, for daily U.S. zero-coupon bond yields at 30 different
maturities from 1 year to 30 (from the J.P. Morgan database), leads to the conclu-
sion that factor loadings are not stable through time. The null hypothesis of equal

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


coefficients for the (adjusted) factor scores in the different subperiods is rejected
in most cases. In particular, the results of our tests indicate that the factor loadings
of the first and second factors are not robust through time at intermediate
maturities (between 8 and 20 years). Furthermore, the loadings of the first factor
in the different subperiods are not robust through time also at long maturities
(longer than 20 years). For short maturities, time instability is also detected, in
particular with respect to the third factor loadings.
Such an instability problem is solved by applying the factor analysis on
appropriate standardized multivariate residuals. Using the same tests introduced
before, we provide empirical evidence that multivariate scaled residuals from our
nonparametric technique display a stable factor structure. This is not the case
when using, for example, a classical parametric VAR-CCC-GARCH model to
filter the interest rate data. This result highlights the strong power and the
accuracy of our filtering technique.
The article is organized as follows. In Section 1, the model used to filter the data
and the estimation algorithm are outlined. The statistic constructed to test the time
robustness hypothesis of the factor loadings is presented in Section 2 and the results
obtained by applying it to the dataset under study are reported in Section 3. Section
4 provides a solution to the instability problem and compares the precision of our
nonparametric filtering techniques with a standard alternative model.

1 FILTERING THE DATA


This section describes the multivariate model used to filter the series of interest
rate changes at different maturities. There are several reasons for using indepen-
dent innovations rather than interest rate levels or changes. The first one is that
interest rate levels are highly cross-correlated. Interest rates at less liquid matu-
rities appear to lag more liquid ones. Therefore part of their change appears to be
predictable from past changes in adjacent rates. Ignoring this fact in the estima-
tion of the covariance matrix leads to a spurious factor structure. The effect of
cross-correlation is that factor analysis on interest rates will consider most of the
information embodied in the term structure as redundant. Hence a single trend
variable will account for most variance. Taking differences reduces the correlation
between interest rates and allows a better study of the factors influencing interest
rate movements. Another equally important reason is that interest rate levels are
highly autocorrelated, nonstationary, and their distribution appears to be slightly
Audrino et al. | The Stability of Factor Models 425

leptokurtic. On the other hand, interest rate changes are stationary, their distribu-
tion is more leptokurtic, with drastically reduced autocorrelations. However,
autocorrelations are not completely removed by differencing. Excess kurtosis
does not affect factor analysis, but time dependence clearly affects estimation
procedures, since one of the assumptions inherent in standard factor analysis is
that the data under consideration represent random, independent samples from a
multivariate distribution. For all these reasons, we filter, in a first step, the
autocorrelation left in the series of interest rate changes using a nonparametric
VAR-type model. In addition, we present a computationally feasible estimation

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


algorithm and a model selection procedure that can be applied to it.

1.1 Starting Point


Our data are time series of (standardized) interest rate changes at different maturities
 
rt ,Ti  rt1 ,Ti
xt ,Ti ¼ ; t ¼ 1,:::, n, i ¼ 1,:::, d , ð1Þ
rt1 ,Ti

where rt,Ti denotes the spot interest rate on day t for maturity Ti. We assume
stationarity of xt,Ti.
Our goal is to filter the autocorrelation of the process {xt} of interest rate
changes by considering the whole term structure dynamics (i.e., looking simulta-
neously at all available maturities), so that a standard factor analysis can be
performed on the filtered data. To this purpose, we consider a nonparametric
VAR-type model in connection with a functional gradient descent (FGD) estima-
tion [see Friedman, Hastie and Tibshirani (2000) and Friedman (2001)] of the
multivariate conditional mean mt ¼ E½xt j F t1  of xt ¼ ðxt,T1 ,:::, xt,Td ÞT , where F t-1
denotes the information available up to time t–1, that is, the s-algebra generated
by fxs ; s  t  1g.
In our model, the dynamics of the conditional mean are specified by
 T
xt ¼ mt þ xt , mt ¼ mt, 1 ,:::, mt ,d ,

mt,i ¼ Gi ðfxtj,Tk ; j ¼ 1, 2,:::; k ¼ 1,:::, dgÞ, i ¼ 1,:::, d; ð2Þ

where we assume that fxt gt is a sequence of nonautocorrelated conditionally


gaussian multivariate innovations having zero mean and covariance matrix
covfxt g ¼ Vt .
The functional form for the conditional means in Equation (2) allows for
crossterms, since the conditional means depend on past multivariate observa-
tions. This is one of the interesting features of the proposed nonparametric VAR-
type model. It is motivated by the consideration that time series of interest rate
changes are strongly cross-correlated because of differences in liquidity. There-
fore most series can be influenced and better predicted using past information
from other series.
As Audrino and Bühlmann (2003) show, an adapted procedure based on FGD
can be a powerful strategy to construct computable estimates and predictions for
426 Journal of Financial Econometrics

the multivariate conditional mean, mt , and volatility matrix, Vt, overcoming the well
known curse of dimensionality problem. All details about the use of FGD for
estimating mt in the general model of Equation (2) are presented in the next section.

1.2 Conditional Mean Estimation Using FGD


We propose FGD to estimate the conditional mean functions Gi ðÞ in Equation (2),
where we restrict Gi ðÞ : Rpd ! R with p finite, that is, involving the first p lagged
multivariate observations. The main idea of FGD is to estimate the functions Gi ðÞ

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


by minimizing a suitable loss function l, under the constraint that the solutions
Ĝi ðÞ are additive expansions of ‘‘simple estimates.’’ These ‘‘simple estimates’’ are
given by a statistical procedure, S, called the base learner, which is often con-
structed from a (constrained or penalized) least squares fitting; common exam-
ples of base learners are regression trees, projection pursuit and neural nets. For
more details, we refer the reader to Friedman, Hastie, and Tibshirani (2000),
Friedman (2001), Audrino and Bühlmann (2003), and Audrino and Trojani (2004).
Analogous to Audrino and Bühlmann (2003), we have that the (multivariate)
negative log-likelihood (conditional on the first p observations) in Equation (2) is
given by
Xn  
d=2 1=2 T 1
 log ð2pÞ detðVt Þ expðxt Vt xt =2Þ , ð3Þ
t¼pþ1

with xt ¼ xt  mt ¼ xt  Gðxt1
tp Þ. For this reason, a natural loss function is

1 1 d
lV ðy,gÞ ¼ ðy  gÞT V 1 ðy  gÞ þ logðdetðVÞÞ þ logð2pÞ,
2 2 2
y  g ¼ ðy1  g1 ,:::, yd  gd ÞT , ð4Þ

where the term dlogð2pÞ=2 is constant and could be dropped. As emphasized by


the subscript, the loss function lV depends on the unknown conditional covar-
iance matrix V. We assume that the conditional covariance matrix dynamics
follow a multivariate constant conditional correlation (CCC) model4 [Bollerslev

4
We assume that

Vt ¼ Dt RDt ,
d ðN1Þ
Dt ¼ diagðst,1, :::,st,d Þ, R ¼ ½rij  i,j¼1:

In this model, the parameter rij ¼ corrðxt;Ti , xt,Tj jF t1 Þ equals the constant conditional correlation and
hence 1  rij  1, rii ¼ 1: Moreover, the individual conditional variances s2t,i are approximated by the
general nonparametric function, Fi,

s2t,i ¼ Fi ðfxtj,Tk ; j ¼ 1, 2,:::; k ¼ 1,:::, dgÞ, i ¼ 1,:::, d,

that are also estimated using the FGD methodology; for all details, see Audrino and Trojani (2004). It
follows that the empirical criterion of Equation (3) and the loss function of Equation (4) can be further
simplified as
Audrino et al. | The Stability of Factor Models 427

(1990)], where the individual conditional variances are also approximated by


some general nonparametric functions Fi and estimated using FGD [for all details,
see Audrino and Trojani (2004)]. Then the FGD algorithm is constructed itera-
tively by estimating V and using the loss function computed with the estimated V,
to estimate all Gi’s. Details of the FGD algorithm are presented in Appendix A.
The filtered innovations are then constructed as

x̂t ¼ xt  m̂t ¼ xt  ĜM ðtÞ, t ¼ p þ 1,:::, n: ð5Þ

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


Note that using the FGD technique, we also eliminate the most important cross-
effects in the conditional mean, caused, for example, by the fact that bonds at
some maturities are more liquid than others. In other words, we filter out from the
data not only the individual serial correlations, but also the most important cross-
terms in the conditional mean. This improves, for example, the model of Stock
and Watson (2002), where, in forecasting, only a diagonal autoregressive term is
considered.
A last remark concerns the consistency of conditional mean and volatility
approximations obtained using the FGD technique. In particular, the assumed
conditional gaussianity of the innovations can be too restrictive for financial data.
FGD is robust against deviations from gaussianity in the following sense. FGD
tries to improve locally some simple parametric starting approximations that are
consistent under pseudo-maximum log-likelihood maximization, achieving some
optimality criterion asymptotically [see, e.g., Bühlmann and Yu (2003)]. There are
several results about consistent function approximation by FGD algorithms
where the only assumption is that the function must have any fixed finite-
dimensional predictor space as it domain. Some references include Zhang
and Yu (2003), Lugosi and Vayatis (2004), and Zhang (2004). Last, but not
least, the good empirical performance of FGD in the financial field [see e.g.,
Audrino and Bühlmann (2003) and Audrino and Trojani (2004)] supports evi-
dence that it is a reliable and robust strategy. This is also apparent from the results
of our study.

2 TEST DESCRIPTION AND IMPLEMENTATION


We perform a robust factor analysis (FA) on filtered innovations, x̂t , obtained
applying the model described in Section 1, which is estimated via FGD. As high-
lighted by Croux et al. (2003) in their study, the use of robust FA takes into account a
possible nongaussianity of the filtered innovations, a feature that is common in

 
n
X 1 T 1
logðdetðDt ÞÞ þ ðD1 1 0 0
t xt Þ R ðDt xt Þ þ n d logð2pÞ=2 þ n logðdetðRÞÞ=2 and
t¼pþ1 2
1 1 T 1 1 1 d
lR ðy, gÞ ¼ logðdetðDÞ þ ðD ðy  gÞÞ R ðD ðy  gÞÞ þ logðdetðRÞÞ þ logð2pÞ,
2 2 2
respectively, where n0 ¼ n  p: Estimation of the correlation matrix R can be easily done via empirical
moments of residuals having (previous) estimates Ĝ ¼ ðĜ1 ,:::, Ĝd Þ and D̂t ¼ diagðŝt,1 ,:::,ŝt,d Þ:
428 Journal of Financial Econometrics

financial data. Alternatively, we also perform tests based on an independent com-


ponent analysis (ICA) of the filtered innovations [see, e.g., Porcaro (2003)]. The
results are qualitatively the same as those presented in the next sections.
The robust FA model with K factors is

X
K
x̂ t ,Ti ¼ LTi, j Ft,j þ Zt, Ti ,
j¼1

where L are the factor loadings, F are the common factor scores, and Zt are the

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


errors (or specific factors). To estimate the optimal parameters in the model we
use the sum over the absolute residuals criterion, that is, the optimal parameters
are those that minimize such a criterion.
Analogous to Perignon and Villa (2002), we divide the period under considera-
tion in four nonoverlapping subperiods of equal length (only the last subperiod has
one extra observation). Let S1 be the first subperiod, that is, S1 ¼ ft : t ¼ 1,:::, n1 g, S2
be the second one, from t ¼ n1 þ 1 to t ¼ n1 þ n2 , and so on. We have four subper-
iods, Sl , with l ¼ 1, 2, 3, 4.
For each maturiy, Ti , we perform a robust linear regression minimizing the
sum of absolute residuals (to be consistent with the criterion used in the robust
FA) using an M-estimator where fitting is done by iterated reweighted least
squares [Hampel et al. (1986)]. In our regressions, the innovation series, x̂t,Ti , is
the dependent variable and the predictors are dummy variables and the vectors of
factors scores multiplied by the corresponding factor loadings, F̂t,j ¼ LTi j Ft,j . Our
dummy variables are constructed as follows:
n
0 if t 2
= Sl
Dlðt,jÞ ¼ F̃t,j o:w:

The linear regression model we estimate is thus

X
K XX
K
x̂t,Ti ¼ ~t,j þ
aj F bl,j Dlðt,jÞ þ Zt, Ti ,
j¼1 l˛D j¼1

where D is one of the sets (2, 3, 4), (1, 3, 4), (1, 2, 4), or (1, 2, 3).
To check the stability over time of factor loadings, we perform a test to verify
the null hypothesis that the coefficients of the dummy variables are equal to zero,
that is, that the regression coefficients (loadings) in the different subperiods are
equal.
To construct the test statistic, we estimate the robust variance-covariance
P
matrix, , following White (1980), by

X
X
n
ˆ ¼ Z2t, Ti BBT ,
t¼pþ1
Audrino et al. | The Stability of Factor Models 429

where B is a t  4K matrix constructed by the combining into a single matrix F̃ and


Dl for l ˛D, so, for example, if D ¼ ð1, 2, 3Þ, then

~ D1 D2 D3 :
B ¼ ½F

The test statistic is obtained by dividing the estimated regression coefficient


pffiffiffiffiffi
of the dummy variable by the corresponding estimated standard deviation, ŝii ,
P
where ŝii is the ith element on the diagonal of ˆ . Upon rejection of the null
hypothesis, we conclude that the corresponding factor is not stable.

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


3 EMPIRICAL RESULTS
3.1 Data and Filtered Innovations
In our empirical investigation, we use daily U.S. zero-coupon bond yields at
d ¼ 30 different maturities from 1 year to 30 years (data are obtained from the
J.P. Morgan database). The time period considered is between January 2, 1986,
and May 10, 1995 (without holidays). Therefore the number of our observations
is n ¼ 2330. In Figure 1, we provide a three-dimensional plot for our term
structure data. Descriptive statistics of our data are available on request.
We focus on the filtered innovations x̂t in Equation (5), computed using the
method presented in Secton 1, which is based on an FGD estimation of the
conditional mean functions. We consider in the estimation only the last past
lagged multivariate observation as predictor,5 that is, p ¼ 1. Three-dimensional
plots and descriptive statistics for the filtered innovations are provided in
Figure 2 and in Table 1, respectively.
As Figure 2 clearly shows, there is an outlier coresponding October 20,
1987. Once it is eliminated, the sample autocorrelations of the filtered innova-
tions summarized in Table 1 are very small, supporting the FGD model
presented in Section 1. Moreover, we compute classical Ljung-Box statistics,
LB(10), testing for autocorrelations in the level of the filtered innovations up
to the 10th order. The test statistics, in 29 of 30 cases, are not significant at the
5% level or better.
In addition, we also perform the generalized Jarque-Bera tests of Chen and
Kuan (2003) for checking departures from normality. These tests are based on the
P1
third and fourth moments of the multivariate standardized residuals ˆ t x̂t , where
P PT
ˆ t ˆ ¼ V̂t in Equation (2). In all 30 cases, the null hypothesis of normality is rejected
t

5
Stock and Watson (2002) present a methodology to forecast a macroeconomic time-series variable where a
large number of predictors is summarized by a small number of indexes constructed by principal
components analysis. An approximate dynamic factor model serves as the statistical framework for the
estimation of the indexes and construction of the forecasts. In their model, factor loadings are allowed to
vary over time and are modeled as evolving according to a random walk. However, they prove uniform
consistency of the factors from the principal components analysis only in the case where the number of
predictors is considerably larger than the number of observations. This is not the case in our study.
430 Journal of Financial Econometrics

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


Figure 1 Three-dimensional plot of the term structure data: the sample consists of 2330 daily
yield data for the time period between January 2, 1986, and May 10, 1995, at all yearly maturities
from 1 to 30 years.

Figure 2 Three-dimensional plot of positive (left) and negative (right) filtered innovations: the
sample consists of 2330 daily yield data for the time period between January 2, 1986, and May 10,
1995, at all yearly maturities from 1 year to 30 years.

at the 5% level. However, the same result occurs when filtering the data with a
standard diagonal VAR model for the conditional mean in Equation (2).
We perform a robust three-factor analysis of the filtered innovations and
investigate the time robustness of factor loadings. Note that applying a robust
Audrino et al. | The Stability of Factor Models 431

Table 1 Summary statistics for filtered innovations computed with the model
presented in Section 1, based on a multivariate FGD estimation of conditional
mean functions, at some significant maturities.
Maturity Mean St. Dev. Min. Max. MAE RMSE r̂ð1Þ r̂ð20Þ r̂ð50Þ

1 year 0.011 1.146 5.758 5.600 0.809 1.145 0 0.069 0.040


2 years 0.011 1.042 5.828 4.915 0.744 1.041 0 0.043 0.023
3 years 0.011 0.990 4.959 4.372 0.712 0.990 0.006 0.027 0.013
5 years 0.012 0.922 3.785 3.766 0.673 0.922 0.009 0.006 0.011

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


7 years 0.013 0.890 4.030 3.679 0.652 0.890 0.015 0.004 0.010
10 years 0.013 0.830 3.580 3.628 0.612 0.830 0.014 0.004 0.016
15 years 0.013 0.805 3.645 3.411 0.589 0.804 0.004 0.008 0.008
20 years 0.012 0.775 3.391 3.469 0.569 0.774 0.002 0.017 0.021
25 years 0.011 0.772 3.484 3.664 0.566 0.772 0.007 0.024 0.028
30 years 0.012 0.785 3.561 3.797 0.574 0.785 0.008 0.029 0.030

The sample consists of 2329 daily yield data for the time period between January 2, 1986, and May 10,
1995, where an outlier corresponding to the crash of the 20th of October 1987 is recognized and eliminated.
St. Dev. and r̂ðxÞ mean sample standard deviation and sample autocorrelation at lag x, respectively. MAE
and RMSE are the mean absolute error and the root mean-squared error of the filtered innovations,
respectively.

FA, we are able to take into account the nongaussianity detected in our filtered
innovations.
We report results for a robust three-factor analysis, because this choice is
popular in the term structure literature. Moreover, testing the optimal number
of factors for our data sample using conventional information criteria and
standard rules, such as the number of eigenvalues greater than one or the
elbow rule for the scree plot, we find that three factors are empirically a good
choice. More in detail, when performing a robust principal component analysis
of the filtered innovations, we find two eigenvalues greater than one. The scree
plot for a robust factor analysis also supports the choice of three factors. In
addition, when setting a 99% threshold for the explained cumulative variance
we choose exactly three factors.
In our empirical analysis6 we set n1 ¼ n2 ¼ n3 ¼ 582 and n4 ¼ 583. Results for
a classical investigation of factor loading stability are summarized in Appendix B.

3.2 Checking the Time Robustness of Factor Loadings Using


Statistical Tests
In Tables 2–4 we show the value of the test statistic, computed as outlined in
Section 2 for three representative maturities: 5 years (representative of short

6
Results do not depend on a particular choice of the subperiods. We obtain similar results also when
dividing the data sample in subperiods of different (not necessarily equal) lengths or considering partial
overlapping subperiods, as in Perignon and Villa (2002).
432 Journal of Financial Econometrics

Table 2 Value of test statistics described in Section 2 for all different D’s and a
representative short maturity of 5 years.
Factor 1 Factor 2 Factor 3

 = {2, 3, 4}
2 – 1 0.1395 0.8252 2.3964
3 – 1 0 0 0
4 – 1 0 0 0

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


 = {1, 3, 4}
1 – 2 0.1395 0.8252 2.3964
3 – 2 0.0896 0.6339 2.3649
4 – 2 0.0872 0.6532 2.6405

 = {1, 2, 4}
1 – 3 0 0 0
2 – 3 0.0896 0.6339 2.3649
4 – 3 0 0 0

 = {1, 2, 3}
1 – 4 0 0 0
2 – 4 0.0872 0.6532 2.6405
3 – 4 0 0 0

Results are reported for a robust three-factor analysis of filtered innovations x̂t using the nonparametric
FGD methodology. One and two stars indicate a rejection of the null hypothesis of equal factor loadings in
the different subperiods at significance levels of 5% and 1%, respectively.

maturities), 15 years (representative of intermediate maturities), and 30 years


(representative of long maturities). We obtain similar results also for all other
short (1–7 years), intermediate (8–20 years), and long (more than 20 years)
maturities.
We present four panels for each representative maturity. Each table uses
a different subset of the four dummy variables, following the order:
D ¼ ð2, 3, 4Þ; ð1, 3, 4Þ; ð1, 2, 4Þ, ð1, 2, 3Þ. In the tables, an asterisk () indicates
that the null hypothesis is rejected at a 5% significant level, while  indicates
rejection at a 1% significance level. Note that zero values in Table 2 indicate
test values that are smaller or equal to 105.
In contrast with the common belief and the empirical resuls found by Perignon
and Villa (2002) and Diebold and Li (2002) for monthly data, we get that, in our daily
dataset, factor loadings are not stable through time. The null hypothesis of equal
coefficients for the (adjusted) factor scores in the different subperiods is rejected in a
large number of cases. In particular, the results of our tests indicate that the factor
loadings of the first and second factors are not constant at intermediate maturities
and the factor loadings of the third factor are not constant at short maturities.
Moreover, loadings of the first factor are also unstable at long maturities.
Audrino et al. | The Stability of Factor Models 433

Table 3 Value of test statistics described in Section 2 for all different D’s and a
representative intermediate maturity of 15 years.
Factor 1 Factor 2 Factor 3

 = {2, 3, 4}
2 – 1 2.7587 0.5829 0.0148
3 – 1 1.0300 3.1976 0.5701
4 – 1 0.3191 3.7888 0.4456

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


 = {1, 3, 4}
1 – 2 2.7587 0.5829 0.0148
3 – 2 2.0640 4.2859 1.1294
4 – 2 4.8858 0.9228 0.7338

 = {1, 2, 4}
1 – 3 1.0300 3.1976 0.5701
2 – 3 2.0640 4.2859 1.1294
4 – 3 1.8936 0.8013 0.4173

 = {1, 2,3}
1 – 4 0.3191 3.7888 0.4456
2 – 4 4.8858 0.9228 0.7338
3 – 4 1.8936 0.8013 0.4173

Results are reported for a robust three-factor analysis of filtered innovations x̂t using the nonparametric
FGD methodology. One and two stars indicate a rejection of the null hypothesis of equal factor loadings in
the different subperiods at significance levels of 5% and 1%, respectively.

How can we interpret these statistical results, also in comparison to the gra-
phical analysis of Appendix B? When performing specific FA in each subperiod, as
in Appendix B, the vector of factor scores spans the space in a different way in each
subperiod. Consequently, factor loadings show similar patterns across different
time periods, even though they are not constant through time. In contrast, in the
tests of Section 3.2, the common factors are estimated from a FA over the whole
period and then fixed. Regressing the filtered innovations on the identified common
factors, we find that factor loadings in the different subperiods are not robust
through time. Therefore a model allowing for time-varying factor loadings should
be used to better understand the behavior of the term structure.
As an alternative, researchers may apply FA on full multivariate scaled
residuals. Clearly, residuals must be obtained by multivariate techniques that are
still computationally feasible in large dimensions. Do multivariate standardized
residuals display a stable factor structure after filtering? In Section 4 we report
results for a parametric VAR-CCC-GARCH model compared with our nonpara-
metric FGD methodology.
To be sure that the factor loadings instability is not due to the selection of an
insufficient number of factors, we also perform the FA for more than three factors.
434 Journal of Financial Econometrics

Table 4 Value of test statistics described in Section 2 for all different D’s and a
representative long maturity of 30 years.
Factor 1 Factor 2 Factor 3

 = {2, 3, 4}
2 – 1 0.9446 1.5402 0.1357
3 – 1 0.6449 0.7930 0.1379
4 – 1 2.0811 0.6954 0.7668

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


 = {1,3,4}
1 – 2 0.9446 1.5402 0.1357
3 – 2 0.3689 0.8659 0.0219
4 – 2 3.9889 4.8587 1.1748

 = {1, 2,4}
1 – 3 0.6449 0.7930 0.1379
2 – 3 0.3689 0.8659 0.0219
4 – 3 4.3956 1.1992 0.6361

 = {1, 2,3}
1 – 4 2.0811 0.6954 0.7668
2 – 4 3.9889 4.8587 1.1748
3 – 4 4.3956 1.1992 0.6361

Results are reported for a robust three-factor analysis of filtered innovations x̂t using the nonparametric
FGD methodology. One and two stars indicate a rejection of the null hypothesis of equal factor loadings in
the different subperiods at significance levels of 5% and 1% respectively.

In particular, a four-factor analysis shows that, as for the case of three factors, the
instability issue remains. In fact, the first, second, and fourth factor loadings are
not constant over time.7

4 TIME ROBUSTNESS OF MULTIVARIATE SCALED RESIDUALS


In this section we present a possible solution to overcome the instability problem
of the latent factor loadings found in Section 3. In particular, we are interested
in verifying whether full multivariate standardized residuals (i.e., filtered inno-
vations in Section 3, standardized using the estimated volatility matrices) from
our nonparametric filtering technique, based on FGD for conditional means and
volatilities, display a stable factor structure. Moreover, we compare the stability of
factor loadings for the multivariate scaled residuals
X X XT
ˆt ¼ ð ˆ t Þ1 x̂t , with t t ¼ Vt ,

7
Explicit results for a four-factor analysis are available on request.
Audrino et al. | The Stability of Factor Models 435

using our nonparametric FGD methodology with a parametric VAR-CCC-


GARCH(1,1) model.
Similar to Section 3, we compute classical Ljung-Box statistics, LB(10), testing
for autocorrelations in the level of the multivariate scaled residuals Êt , using the
two alternative approaches up to the 10th order. Both models work quite well,
with a little advantage of the nonparametric FGD methodology over the full
parametric VAR-CCC-GARCH(1,1) model. In particular, the test statistics are
not significant at the 5% level in about 85% and 82% of the cases, respectively.
The values of the test statistic introduced in Section 2, for the same repre-

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


sentative maturities of Section 3.2, are summarized in Table 5–7. Results are
reported for a robust three-factor analysis of the multivariate scaled residuals ˆt
from our nonparametric FGD methodology (FGD) and a standard parametric
VAR-CCC-GARCH(1,1) fit (VAR).
As before, we present four panels for each representative maturity. Each table
uses a different subset of the four dummy variables, following the same order
used in the previous tables.

Table 5 Value of test statistics described in Section 2 for all different D’s and a
representative short maturity of 5 years.
Factor 1 Factor 2 Factor 3

FGD VAR FGD VAR FGD VAR

D = {2, 3, 4}
2 – 1 0 0.2958 0 0.7915 0 1.8174
3 – 1 0 0.3471 0 0.6410 0 1.6644
4 – 1 0 0.9461 0 2.2191 0 0.6721

D = {1, 3, 4}
1 – 2 0 0.2958 0 0.7915 0 1.8174
3 – 2 0 0.1178 0 1.6551 0 1.1788
4 – 2 0 1.2142 0 1.7965 0 1.2411

D = {1, 2, 4}
1 – 3 0 0.3471 0 0.6410 0 1.6644
2 – 3 0 0.1178 0 1.6551 0 0.1788
4 – 3 0 1.1244 0 3.0025 0 0.9792

D = {1, 2, 3}
1 – 4 0 0.9461 0 2.2191 0 0.6721
2 – 4 0 1.2142 0 1.7965 0 1.2411
3 – 4 0 1.1244 0 3.0025 0 0.9792

Results are reported for a robust three-factor analysis of multivariate scaled residuals x̂t using the
nonparametric FGD methodology (FGD) and a standard parametric VAR-CCC-GARCH (1,1) model
(VAR). One and two asterisks indicate a rejection of the null hypothesis of equal factor loadings in the
different subperiods at significance levels of 5% and 1%, respectively.
436 Journal of Financial Econometrics

Table 6 Value of test statistics described in Section 2 for all different D’s and a
representative intermediate maturity of 15 years.
Factor 1 Factor 2 Factor 3

FGD VAR FGD VAR FGD VAR

D = {2, 3, 4}
2 – 1 1.0359 2.1252 1.5813 0.3351 0.7784 1.7290
3 – 1 0.1015 1.5042 0.7657 0.1866 0.3809 2.3779

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


4 – 1 0.4527 0.0517 0.0189 1.4746 0.4404 0.6721

D = {1, 3, 4}
1 – 2 1.0359 2.1252 1.5813 0.3351 0.7784 1.7290
3 – 2 2.1852 1.1151 1.2393 0.3903 1.3451 0.8059
4 – 2 1.1948 3.1582 1.5786 0.9014 0.5183 0.4138

D = {1, 2, 4}
1 – 3 0.1015 1.5042 0.7657 0.1866 0.3809 2.3779
2 – 3 2.1852 1.1151 1.2393 0.3903 1.3451 0.8059
4 – 3 1.1307 2.4221 0.8848 0.5795 1.2613 1.2891

D = {1, 2, 3}
1 – 4 0.4527 0.0517 0.0189 1.4746 0.4404 0.6721
2 – 4 1.1948 3.1582 1.5786 0.9014 0.5183 0.4138
3 – 4 1.1307 2.4221 0.8848 0.5795 1.2613 1.2891

ˆ t using the
Results are reported for a robust three-factor analysis of multivariate scaled residuals ˛
nonparametric FGD methodology (FGD) and a standard parametric VAR-CCC-GARCH (1,1) model
(VAR). One and two asterisks indicate a rejection of the null hypothesis of equal factor loadings in the
different subperiods at significance levels of 5% and 1%, respectively.

Tables 5–7 clearly show the strong potential of the nonparametric FGD
technique, used as a filter for the raw interest rate data. In fact, the multivariate
scaled residuals from the FGD fit display a stable factor structure. On no occasion we
reject the null hypothesis that the coefficients of the dummy variables are equal to
zero at the 1% confidence level and only on one occasion at the 5% confidence level.
In contrast, filtering the data using a full parametric VAR-CCC-GARCH model yields
several rejections of the null hypothesis at both the 5% and 1% confidence levels for
short and intermediate maturities. When considering long maturities, the factor
decomposition of the multivariate scaled residuals is stable for both approaches.
This result highlights even more the fact that a stable factor structure of the
interest rate data connot be generally assumed, but it can be obtained if an
appropriate multivariate filtering technique is implemented. It follows the need
of filtering the raw data using nonparametric methodologies that are able to
eliminate all important linear and nonlinear cross-dependencies among the dif-
ferent series. A simple parametric model is not able to perform as well as our
nonparametric model based on FGD. Furthermore, results of our tests indicate
Audrino et al. | The Stability of Factor Models 437

Table 7 Value of test statistics described in Section 2 for all different D’s and a
representative long maturity of 30 years.
Factor 1 Factor 2 Factor 3

FGD VAR FGD VAR FGD VAR

D = {2, 3, 4}
2 – 1 0.7421 0.9155 0.1130 0.1912 0.2512 0.3432
3 – 1 0.1411 0 0.6158 0 0.7694 0

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


4 – 1 1.0993 0 0.9074 0 0.7894 0

D = {1, 3, 4}
1 – 2 0.7421 0.9155 0.1130 0.1912 0.2512 0.3432
3 – 2 0.6542 0.6509 0.4649 0.1881 0.4516 0.2989
4 – 2 0.3167 0.8101 0.6936 0.1996 0.4522 0.3476

D = {1, 2, 4}
1 – 3 0.1411 0 0.6158 0 0.7694 0
2 – 3 0.6542 0.6509 0.4649 0.1881 0.4516 0.2989
4 – 3 0.8915 0 0.1587 0 0.0175 0

D = {1, 2, 3}
1 – 4 1.0993 0 0.9074 0 0.7894 0
2 – 4 0.3167 0.8101 0.6936 0.1996 0.4522 0.3476
b3 – b4 0.8915 0 0.1587 0 0.0175 0

ˆ t using the
Results are reported for a robust three-factor analysis of multivariate scaled residuals ˛
nonparametric FGD methodology (FGD) and a standard parametric VAR-CCC-GARCH (1,1) model
(VAR). One and two asterisks indicate a rejection of the null hypothesis of equal factor loadings in the
different subperiods at significance levels of 5% and 1%, respectively

how the detected instability problem of the factor structure of interest rate data
can be overcome. It is necessary to consider full nonparametric techniques, like
the FGD one, to filer raw interest rate data. Multivariate scaled residuals ensure a
better consistency than simple innovations, filtered only to eliminate cross-corre-
lations and autocorrelations in their conditional means.

5 CONCLUSION
Filtering interest rates so that innovations conform to standard assumptions in
multivariate statistics has become possible. Our use of FGD produces innovations
that are well described by three factors with stable loadings. We develop and
perform tests of the stability of factor loadings. Results support our parsimonious
description of term structure innovations, which appear to be driven by three
variables. On the contrary, our results also show that a standard parametric VAR-
CCC-GARCH model is not sufficiently precise to filter raw interest rate data and
does not yield multivariate scaled residuals displaying a stable factor structure.
438 Journal of Financial Econometrics

APPENDIX A: ESTIMATING CONDITIONAL MEANS


Step 1 (initialization). Choose an appropriate starting function Ĝi,0 ðÞ:8 Denote by

Ĝi,0 ðtÞ ¼ Ĝi,0 ðxt1 , xt2 ,:::Þ, i ¼ 1,:::, d:


ð0Þ
Estimate V̂t as in Equation (N1) using Ĝ0 . Set m ¼ 1.
For every component i ¼ 1,:::, d, do the following.
Step 2i (projection of component gradients to base learner). Compute the negative gradient:

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


ql ðm1Þ ðxt, GÞ
V̂t
Ut,i ¼  jG¼Ĝm1 ðtÞ , t ¼ p þ 1,:::, n:
qGi
Then, fit the negative gradient vector Ui ¼ ðUpþ1,i,:::, Un,i ÞT with a statistical proce-
dure called base learner, always using the first p time-lagged predictor variables
(i.e., xt1
tp is the predictor for Ut,i )

ĝm,i ðÞ ¼ SX ðUi ÞðÞ,

where SX ðUi ÞðxÞ denotes the predicted value at x from the base learner S using the
response vector Ui predictor variables X.9
Step 3i (line search). Perform a one-dimensional optimization for the step length,
X
n
ŵm,i ¼ argminw l ðm1Þ ðxt, Ĝm1 ðtÞ þ wĝm,i ðxt1
tp ÞÞ,
V̂t
t¼pþ1

where Ĝm1 ðtÞ þ wĝm,i ðÞ is defined as the function which is constructed by
adding in the ith component only. (Note that the line search guarantees that the
negative log-likelihood decreases in every iteration.)
Step 4 (update). Select the best component as
X
n 
im ¼ argmini l ðm1Þ ðxt, Ĝm1 ðtÞ þ ŵm,i ĝm,i xt1
tp :
V̂t
t¼pþ1

Update10

8
Initialization is important to achieve good estimates. As a starting function we propose to use the fit from
a parametric diagonal VAR (1)-CCC-GARCH(1,1) model.
9
The base learner clearly determines the FGD estimates ĜM ðÞ: This should be ‘‘weak’’ (not involving too
many parameters to be estimated) enough not to immediately produce an over fitted estimate at the first
iteration. The complexity of the FGD estimates ĜM ðÞ is increased by adding further terms at every
iteration. We choose decision trees as base learners, since, particularly in high dimensions, they have the
ability to do variable selection by choosing few of the explanatory variables for prediction.
10
It is often desirable to make a base learner sufficiently ‘‘weak.’’ A simple, but effective way to reduce the
complexity of the base learner is via shrinkage toward zero. The update is then replaced by

Ĝm ðÞ ¼ Ĝm1 ðÞ þ n  ŵm,im ĝm,im ðÞ: ðN2Þ

Obviously this reduces the variance of the base learner by the factor n.
Audrino et al. | The Stability of Factor Models 439

Ĝm ðÞ ¼ Ĝm1 ðÞ þ ŵm,im ĝm,im ðÞ:


ðmÞ
Then compute the new estimates V̂t according to Equation (N1) using Ĝm .
Step 5 (iteration). Increase m by one and iterate Steps 2–4 until stopping with
m ¼ M. This produces the FGD estimates for conditional means,
X
M
ĜM ðÞ ¼ Ĝ0 ðÞ þ ŵm,im ĝm,im ðÞ:
m¼1

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


The stopping rule in Step 4 is required to avoid overfitting. Our stopping criterion
can be viewed as a regularization device that is very effective in complex model
fitting. We choose the stopping value M with the following scheme: we split the
(in-sample) estimation period into two sets, the first of size 0:7  n used as a
training set and the remaining period used as a test set (this stopping procedure
can also be adopted when the data are dependent). The optimal value of M is then
chosen to optimize the cross-validated log-likelihood. Empirical studies by
Audrino and Bühlmann (2003) show that estimating M by a simple 70%–30%
cross-validation scheme works well.

APPENDIX B: CHECKING THE TIME ROBUSTNESS OF FACTOR


LOADINGS IN THE CLASSICAL WAY
We test graphically the time robustness of factor loadings, as in Bliss (1997). To
this purpose, for each subperiod we perform a three-factor analysis of the filtered
innovations, x̂t . The resulting rotated11 factor loadings for each of the subperiods
are shown in Figure 3.
The first factor is very close to constant across all maturities. This result is true
for all four subperiods. Therefore a change in the first factor will produce a
parallel movement in all interest rates and factor 1 can be interpreted as a level
factor as usually.
The loadings on the second factor decrease uniformly from a moderate positive
value at short maturities, to a moderate negative value at long maturities. Once
again, the shapes and the values of the short and long ends of these curves are
approximately the same across subperiods. Factor 2 produces movements in the
long and short ends of the curve in opposite directions (twisting the curve), with
commensurate smaller changes at intermediate maturities.
As expected, factor 3 may be interpreted as a curvature factor. The
loadings are zero for the shortest maturities, indicating the short rates are
unaffected by factor 3, positive for intermediate maturities and then decline to
become negative for the longest maturities. Once again the shape and the

11
Similar to Bliss (1997), we rotate the original solution until factor loadings with a meaningful economic
interpretation are obtained. For instance, for this article the loadings were first rotated so that filtered
innovations at all maturities had approximately the same loading on the first factor. Since any
perturbation to the first factor then affects all maturities equally, this factor is interpreted as a level
shift.
440 Journal of Financial Econometrics

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


Figure 3 Rotated factor loadings of the first (top), second (center), and third factor (bottom).
Results are obtained by splitting the sample in four subperiods and performing a different FA for
each subperiod. Subperiod-specific factor loadings are plotted with different line types (solid,
dotted, dashed, and long dashed). The whole sample consists of 2329 daily yield data for the time
period between January 2, 1986, and May 10, 1995, at all yearly maturities from 1 year to 30 years.
The first subperiod consists of the first n1 observations, the second of the following n2 observa-
tions, and so on. For our analysis we take n1 ¼ n2 ¼ n3 ¼ 582 and n4 ¼ 583:

values of the curves in the four subperiods are approximately the same, even
if in the first subperiod the peak is reached after those of the other
subperiods.
Summarizing, from a graphical point of view, we find that factor loadings
seem to present a consistent pattern across different subperiods despite some
minor variation. This is similar to what is reported by Bliss (1997).

Received March 3, 2004; revised November 29, 2004; accepted March 8, 2005.

REFERENCES
Audrino, F., and P. Bühlmann. (2003). ‘‘Volatility Estimation with Functional Gradient
Descent for Very High-Dimensional Financial Time Series.’’ Journal of Computational
Finance 6, 1–26.
Audrino, F., and F. Trojani. (2004). ‘‘Accurate Yield Curve Scenario Generation Using
Functional Gradient Descent.’’ Unpublished manuscript, USI, Lugano, Switzerland.
Bliss, R. R. (1997). ‘‘Movements in the Term Structure of Interest Rates.’’Federal Reserve
Bank of Atlanta Economic Review 82, 16–33.
Audrino et al. | The Stability of Factor Models 441

Bollerslev, T. (1990). ‘‘Modelling the Coherence in Short-Run Nominal Exchange Rates:


A Multivariate Generalized ARCH Model.’’ Review of Economics and Statistics 72,
498–505.
Bühlmann, P., and B. Yu. (2003). ‘‘Boosting with the L2 Loss: Regression and
Classification.’’ Journal of the American Statistical Association 98, 324–339.
Carvelhill, A., and C. Strickland. (1992). ‘‘Money Market Term Structure Dynamics
and the Volatility Expectations,’’ working paper, Financial Options Research
Centre, University of Warwick.
Chapmen, D. A., and N. D. Pearson. (2001). ‘‘Recent Advances in Estimating

Downloaded from http://jfec.oxfordjournals.org/ at Cambridge University Library on March 19, 2013


Term-Structure Models.’’ Financial Analysts Journal 57, 77–95.
ChenY. –T., and C.-M. Kuan. (2003). ‘‘A Generalized Jarque-Bera Test of Conditional
Normality,’’ working paper, Academia Sinica.
Croux, C., P. Filzmoser, G. Pison, and P. J. Rousseeuw. (2003). ‘‘Robust Factor Analysis.’’
Journal of Multivariate Analysis 84, 145–172.
Diebold, F. X., and C. Li. (2002). ‘‘Forecasting the Term Structure of Government Bond
Yield.’’ Forthcoming in the Journal of Econometrics.
Friedman, J. H. (2001). ‘‘Greedy Function Approximation: A Gradient Boosting
Machine.’’ Annals of Statistics 29, 1189–1232.
Friedman, J. H., T. Hastie, and R. Tibshirani. (2000). ‘‘Additive Logistic Regression:
A Statistical View of Boosting. Annals of Statistics 28, 337–407 (with discussion).
Hampel, F. R., E. M., Ronchetti, P. J. Rousseeuw, and W. A. Stahel. (1986). Robust
Statistics: The Approach Based on Influence Functions . New York: John Wiley.
Litterman, R., and J. Scheinkman. (1991). ‘‘Common Factors Affecting Bond Returns’’.
Journal of Fixed Income 1, 54–61.
Leikkos, I. (2000). ‘‘A Critique of Factor Analysis of Interest Rates’’ Journal of Derivatives
8, 72–83.
Lugosi, G., and N. Vayatis. (2004). ‘‘On the Bayes-Risk Consistency of Regularized
Boosting Methods.’’ Annals of Statistics 32, 30–55.
Perignon, C., and C. Villa. (2002). ‘‘Permanent and Transitory Factors Affecting the
Dynamics of the Term Structure of Interest Rates,’’ working paper, FAME.
Perignon, C., and C. Villa (2004). ‘‘Sources of Time Variation in the Covariance Matrix
of Interest Rates.’’ Forthcoming in the Journal of Business.
Phoa, W. (2000). ‘‘Yield Curve Risk Factors: Domestic and Global Contexts.’’ In
L. Borodovsky and M. Lore, eds., The Practitioner’s Handbook of Financial Risk
Management . Woburn, MA: Butterworth-Heinemann.
Porcaro, C. (2003). ‘‘Applying Independent Component Analysis to Factor Model in
Finance.’’ Master’s thesis, Università degli Studi di Roma ‘‘La Sapienza.’’
Stambaugh, R. F. (1988). ‘‘The Information in Forward Rates: Implications for Models
of the Term Structure.’’ Journal of Financial Economics 21, 41–70.
Steeley, J. M. (1990). ‘‘Modelling the Dynamics of the Term Structure of Interest Rates.’’
Economic and Social Review-Symposium on Finance 21, 337–361.
Stock, J. H., and M. W. Watson. (2002). ‘‘Macroeconomic Forecasting Using Diffusion
Indexes.’’ Journal of Business and Economic Statistics 20, 147–162.
White, H. (1980). A Heteroskedasticity-Consistent Covariance Matrix Estimator and a
Direct Test for Heteroskedasticity. Econometrica 48, 817–838.
Zhang, T. (2004). ‘‘Statistical Behavior and Consistency of Classification Methods
Based on Convex Risk Minimization.’’ Annals of Statistics 31, 56–134.
Zhang, T., and B. Yu. (2003). Boosting with Early Stopping: Convergence and
Consistency. Forthcoming in the Annals of Statistics.

You might also like