Bispap65o RH
Bispap65o RH
Abstract
We present evidence on the changing dynamics of the yield curve from 1998 to 2011. We
identify four different phases. As expected, the financial crisis represents a period of elevated
yield volatility, but it can be split into two distinct periods. The split occurs when the Federal
Reserve reached the zero lower bound. This bound suppressed volatility in the short end of
the yield curve while increasing volatility in the long end – despite lower overall volatility in
financial markets. In line with previous studies, we find that announcements with regard to
the Federal Reserve’s large scale asset purchases reduce longer term yields. We also
quantify the effect of widely observed economic news, such as the non-farm payrolls and
other items, on the yield curve.
Keywords: Term structure of interest rates, financial crisis, interest rate dynamics, LSAP,
unconventional monetary policy
JEL classification: E43, E52
1
We thank Torsten Ehlers, Jacob Gyntelberg and Peter Kugler for helpful discussions and we thank Andrew
Filardo for his useful comments. The views expressed in this paper are those of the authors and do not
necessarily reflect those of the Bank for International Settlements.
2
Bank for International Settlements.
3
Dept. of Economics, University of Basel, Peter Merian-Weg 6, P.O. Box, CH–4002 Basel, Switzerland.
Corresponding author. Email addresses: [email protected] (Morten L. Bech), [email protected]
(Yvan Lengwiler).
2. Intelligible factors
We use the decomposition of the term structure into “intelligible factors” developed by
Lengwiler and Lenz (2010). We have M maturities that we observe on T days. Let rt m
denote the interest rate for a zero bond at time t which matures at time t m . The cross
section of interest rates is described by three factors,
rt m k1 m 1,t k2 m 2,t k3 m 3,t t m , (1)
where k M 3 matrix are the loadings and 3 T matrix are time-varying factors.
and k are constructed together so that they have certain desirable properties. Firstly,
constraints are imposed on the loadings k , such that they load on different parts of the
maturity spectrum, as can be seen from Figure 1. The first factor is the only one that loads on
the very long end of the maturity spectrum, so we call 1 the long factor. The second factor is
the only one that loads on the very short end of the maturity spectrum, so we call 2 the short
factor. The third factor has zero loading at the short and the long end of the maturity
Figure 1
Loadings of the three factors
4
Note that these loadings differ from the more common loadings “level”, “slope”, and “curvature”, which
have become custom in applications of principal component analysis (Litterman and Scheinkman, 1991)
or in the specification of Nelson and Siegel (1987).
Figure 2
Estimated intelligible factors
This stylized fact is also true in the extended data sample. Figure 3 depicts the variance
decomposition, i.e. the parts of the variance of the interest rates that are due to the innovations
into the three factors, u . The variance decomposition firstly reveals that the model captures
the second moment of the yields – the term structure of interest rate variance – very well, and
secondly confirms that most of the yield curve movements have their source in innovations into
the curvature factor.
5
See www.treasury.gov/resource-center/data-chart-center/interest-rates/Pages/yieldmethod.aspx for a description
of the methodology.
6
“Local volatility” is a non-parametric measure of the second moment. It is essentially a Nadaraya-Watson
kernel regression on the squared innovations; details are explained in Appendix A.
Table 1
Significant shocks to the yield curve during the four phases
Theoretical normal great liquidity zero lower
phase moderation crisis bound
# cases share # cases share # cases share # cases share
long only 4.51% 49 2.98% 5 0.66% 144 14% 77 10.6%
short only 4.51% 30 1.82% 4 0.53% 50 14.8% 0 0.00%
curv only 4.51% 57 3.46% 10 1.32% 36 10.7% 4 0.55%
long & short only 0.24% 1 0.06% 0 0.00% 5 1.48% 0 0.00%
long & curv only 0.24% 9 0.55% 0 0.00% 6 1.78% 12 1.65%
short & curv only 0.24% 10 0.61% 0 0.00% 19 5.62% 2 0.27%
all three 0.01% 2 0.12% 0 0.00% 5 1.48% 0 0.00%
The shift of the location of the innovations during the financial crisis, and in particular to the
longer part of the maturity spectrum when the zero lower bound became binding, also
manifests itself in the variance attribution. We compute the variance of the yields separately for
the four phases; see Figure 6 and compare with Figure 3 for the whole sample. The differences
are striking. First of all, the overall variance of the yields has dramatically decreased for shorter
maturities in the “zero lower bound phase”. This is a direct corollary of the fact that the zero
lower bound does not allow rates to decrease further, and the Federal Reserve has not allowed
the short rates to increase, hence volatility in this duration spectrum has vanished. As a result,
all the volatility that remains is at longer maturities. The volatilities of the ten- and twenty-year
yields is more or less unchanged for the two subperiods. Yet, because no further movements
at the short end are possible, and the major innovations now occur in the long factor, almost all
of the term structure of yield variance can be attributed to long innovations.
Figure 6
Variance of yields of different maturities and shares
explained by the three types of innovations, in the four phases
Tables 2, 3, and 4 collect the twenty-five largest innovations (in absolute terms) for the long,
short, and curvature factor, respectively, and also report potentially related economic or
financial events. To better gauge the size of these innovations we also divide them, in the
seventh column, by the unconditional standard deviation, by the local volatility estimate for
7
http://timeline.stlouisfed.org.
8
http://www.treasury.gov/resource-center/data-chart-center/quarterly-refunding/Pages/default.aspx.
Table 2
Twenty-five most important innovations to the long factor. The second column reports the size of the innovation
in basis points (bps), and, in parentheses, relative to the unconditional and the local volatility of that day,
respectively. For instance, the largest absolute long innovation is measured on March 18, 2009. We measure a
–53 bps shock. This is 8.0 times the unconditional standard deviation of the long innovation series, and it is
12.3 times larger than the local volatility of that day.
Twenty-five most important innovations to the short factor; see Table 2 for explanation.
Twenty-five most important innovations to the curvature factor; see Table 2 for explanation.
Overall, we note that the larger volatility of long factor innovations in the “zero lower bound”
phase is only partly due to announcements concerning unconventional monetary policy
measures. Of the twenty-five events reported in Table 2, thirteen occurred in the “zero lower
bound” phase. Only four of these large innovations are related to announcements of the Fed
concerning unconventional monetary policy. These are the enlargement of QE 1 (March 18,
2009, –53 bps), the confirmation of the reinvestment policy (December 14, 2010, +25 bps),
one day after Operation Twist 2 was announced (September 22, 2011, –23 bps), and finally
the “forward guidance” announcement (August 9, 2011, –22 bps). All other major shocks
Figure 7
Major long factor innovations in the zero lower bound phase. The last five months of the sample are depicted
separately in the right-hand panel because there is more action there.
5. News
It is well-documented that economic news releases and in particular surprises from market
expectations move Treasury yields (see e.g. Fleming and Remolona, 1999). Not surprisingly,
a similar relationship holds for our factors. A natural question is whether or not the changing
dynamics of the yield curve can (in part) be explained by changing dynamics in terms of
economic news surprises. Table 5 shows the results of regressing innovations into our long,
short, and curvature factors on day-of-release surprises for a range of economic indicators.
The indicators we consider include the Conference Board Consumer Confidence Index®, the
Institute for Supply Management (ISM) purchasing managers index (PMI), the advance GDP
print, the unemployment rate, industrial production, retail sales, housing starts, one-family
houses sold, the personal consumption expenditure price index, capacity utilization, initial
jobless claims, the leading economic indicator index, and the federal funds rate target. We
measure surprises as the difference between the actual value released and the median value
from an “expectations” survey among Wall Street economists conducted by Bloomberg News
prior to the release. To put the surprises on a common scale we standardized them by their
standard deviation over the sample. Moreover, we switch the sign of some surprise
variables, so that a positive surprise is “good news”. For instance, non-farm payroll surprises
are measured as actual release minus median expectation, whereas jobless claims are
defined the other way around. In addition, we control for non-linear effects by including the
squared standardized surprises as additional regressors.
Besides the results for the entire sample, we also split our sample in two with a view to
investigating whether or not the impact of economic news has changed with the financial
crisis. Our “pre-crisis” sample runs from 1998 to August 8, 2007 (“normal” and “great
moderation” phases) and the “crisis” sample covers the remainder of our sample (“liquidity
6. Conclusions
The financial crisis has deeply affected financial markets as well as the economy as a whole.
This has also affected the yield curve and its dynamics. We document these changes in this
paper. Our main results can be summarized as follows: Firstly, we divide the dynamics of the
yield curve into four phases. The first two phases occur prior to the financial crisis. The
second phase is characterized by substantially less volatility of the yields compared to the
first phase. However, this is well explained by the simultaneous decline of overall financial
market volatility during that period as measured by the VIX.
The third and the fourth phase comprise the financial crisis. This means that we can divide
the crisis into two distinct subperiods. In the first subperiod, the yield curve experienced very
strong shocks in the short and medium maturity spectrum due to the freezing of the money
market and subsequent emergency measures taken by the Federal Reserve. The second
part of the financial crisis began when the federal funds rate reached the zero lower bound.
From that point forward, we find an absence of shocks hitting the yield curve at low and
medium maturities. Instead, the longer end of the curve experiences greater disturbances
than before.
Secondly, we perform a (reverse) event study in which we match the greatest shocks to the
yield curve with headline news. We find that some large shocks are associated with
Table 5
Effects of macroeconomic news on the yield curve
sample pre-crisis crisis all pre-crisis crisis all pre-crisis crisis all
constant –0.033 –0.090 –0.065 0.235 0.530 0.328 0.206 –0.195 0.122
standardized surprise in:
Capacity
utilization 2.205** 0.813 2.186** –0.740 –0.457 –0.575 1.233** –0.224 0.808**
Consumer
confidence 1.350** –0.311 1.135** –0.358 0.191 –0.135 0.449 0.224 0.360
Initial jobless
claims 0.975 2.955** 1.620** –0.027 –0.160 –0.174 1.005** –0.553 0.661**
Federal funds
target –2.689 –7.409 –1.334 –12.83*** –1.362 –10.40*** –0.471 3.792 –1.001
Advance GDP –0.668 –2.094 –0.809 0.018 0.697 –0.038 0.528 –0.654 0.571
One-family
houses sold 0.996** 0.711 0.999** –0.129 –0.076 –0.102 –0.047 0.502 0.010
Housing starts –0.407 2.167 –0.179 –0.103 –1.067 –0.264 0.055 1.795* 0.126
Industrial
production –0.990 2.364 –0.694 0.439 0.098 0.334 –0.480 0.537 –0.199
ISM PMI 1.597*** 3.136** 1.941*** –1.305*** 0.348 –0.483* 1.370*** –0.620 0.888***
LEI 0.274 0.175 0.175 –0.833 –0.287 –0.502 –0.708 –0.574 –0.754**
Non-farm
payrolls 3.578*** 4.786** 3.761*** –2.009*** –2.390*** –2.206*** 3.117*** 1.938** 2.948***
PCE price
index 0.356 –3.064* –0.482 –1.324* –0.351 –0.724 –0.141 –0.345 –0.124
Retail sales 1.445** 5.530*** 2.110*** –0.570 0.933 –0.143 1.176*** –1.490* 0.581*
Unemployment
rate 0.783 1.477 0.961* –0.542 –0.524 –0.644** 2.065*** 0.837* 1.477***
1 / 1 0.077*** 0.092 0.098*** 0.260*** 0.271*** 0.298*** 0.144*** 0.251*** 0.178***
1 / 1 0.013 0.034 0.012 –0.061*** –0.112*** –0.072*** –0.024 –0.003 –0.021
log 1
2
0.969*** –0.022 0.952*** 0.931*** 0.887*** 0.907*** 0.965*** 0.891*** 0.956***
great moderation
dummy 0.002 –0.006 –0.004 –0.039* 0.000 –0.014
liquidity crisis
dummy 0.015* 0.145*** 0.051***
zero lower
bound dummy 0.415** 0.026** –0.288*** –0.086*** –0.166** –0.010
VIX 0.002*** 0.043** 0.002*** 0.004*** 0.001 0.003** 0.002** 0.002 0.001**
2
R 0.045 0.032 0.035 0.015 0.003 0.007 0.066 0.009 0.033
Note: *,**,*** denotes significance at the 10%, 5%, and 1% level, respectively.
This is a purely technical appendix which is not necessary to understand the economic
content of the paper. It explains the concept of “local volatility” that is used in some places in
the main part of the article.
We aim at quantifying the changing volatility of the innovations u . Simple visual inspection
suggests heteroscedasticity. We fully acknowledge that this feature of the data is not in line
with the specification of the model. After all, we assumed normally distributed homoscedastic
innovations when estimating the loadings and the VAR with maximum likelihood. Taking the
heteroscedasticity fully into account at the estimation stage of the model seems very
challenging. Being aware of this inconsistency, here we simply aim to measure the stochastic
volatility of the innovations as they present themselves from the model that was estimated
assuming homoscedasticity.
Consider a continuous-time diffusion,
dX t t dt t dWt , (A.1)
where Wt is a standard Brownian motion. t2 is the spot variance process, which is not
observed. Instead, we observe only X t at discrete points in time, t1 t 2 ... t n . Based on
2
work by Bandi and Phillips (2003), Kristensen (2010) establishes that one can estimate
as
K h t i 1 X t i X ti 1
n 2
ˆ
2
i 1
, (A.2)
K h t i 1
T
i 1
In order to estimate spot volatility, two choices need to be made, namely the specification of
the kernel function K and the selection of the bandwidth h . To select the bandwidth, we use
the cross-validation technique; that is, we minimize the mean squared “leave-one-out”
residuals. The kernel function K is symmetric around zero if it weighs observations in the
future the same way as observations in the past. Most popular kernel functions have this
property. Symmetric kernel functions have, however, the disadvantage that for close to the
edge of the sample, they assign positive weights to observations outside the available
sample, which biases the estimation. This is a well-known problem in non-parametric
econometrics.
One way to address this problem is to use a locally adapting kernel function. Such a function
was for instance proposed by Brown and Chen (1999) and Chen (2000). Their kernel
function, based on the beta-function, automatically adapts to the boundaries of the sample:
for close to the first observation t1 , the kernel is right-sided, for close to the last
observation t n , the kernel is left-sided. We have experimented with this kernel but found it to
give unsatisfactory estimates in our application. The volatility measure has significantly more
high-frequency variability close to the edge of the sample than in the interior, which suggests
that the precision of the estimate deteriorates close to the edge, or that the bandwidth
becomes too small.
Because the spot volatility does not appear reasonable, we compute a slightly simpler and
maybe more transparent measure. We apply the Nadaraya-Watson kernel regression on the
squared innovations directly instead of on the squared first differences,
K t X
n 2
i 1
2 h i ti
. (A.3)
K t
T
i 1 h i
This is the same as the approach suggested by Carroll (1982) and Hall and Carroll (1989).
They consider a model where the mean can be parametrically estimated but the variance
cannot. In our case, t is zero by definition, so the setting is simpler.
Figure A.8
Curvature innovations and 50% confidence interval using local volatility estimate
We use the same kernel function and reflection technique as before and perform the cross-
validation bandwidth optimization. The result is an estimate of the volatility that seems much
more reasonable. We call the square root of 2 the local volatility, in order to distinguish it
References
Bai, J., Perron, P., 1998. Estimating and testing linear models with multiple structural
changes. Econometrica 66, 47–78.
Bai, J., Perron, P., 2003. Computation and analysis of multiple structural change models.
Journal of Applied Econometrics 18, 1–22.
Bandi, F., Phillips, P., 2003. Fully nonparametric estimation of scalar diffusion models.
Econometrica 71, 241–283.
Brown, B., Chen, S., 1999. Beta-Bernstein smoothing for regression curves with compact
support. Scandinavian Journal of Statistics 26, 47–59.
Carroll, R., 1982. Adapting for heteroscedasticity in linear models. The Annals of
Statistics 10, 1224–1233.
Chen, S., 2000. Beta kernel smoothers for regression curves. Statistica Sinica 10, 73–91.
Fleming, M., Remolona, E., 1999. Price formation and liquidity in the U.S. Treasury market:
The response to public information. Journal of Finance 54, 1901–1915.
Hall, P., Carroll, R., 1989. Variance function estimation in regression: The effect of estimating
the mean. Journal of the Royal Statistical Society, Series B 51, 3–14.
Kristensen, D., 2010. Nonparametric filtering of the realized spot volatility: A kernel-based
approach. Econometric Theory 26, 60–93.
Lengwiler, Y., Lenz, C., 2010. Intelligible factors for the yield curve. Journal of
Econometrics 157, 481–491.
Litterman, R., Scheinkman, J., 1991. Common factors affecting bond returns. The Journal of
Fixed Income 1, 54–61.
Nelson, C., Siegel, A., 1987. Parsimonious modeling of yield curves. The Journal of
Business 60, 473–489.
Schuster, E., 1991. Incorporating support constraints into nonparametric estimators of
densities. Communications in Statistics: Theory and Methods 14, 1123–1136.