Lecture 3: Forecasting interest

rates (Part 1)

Prof. Massimo Guidolin

Advanced Financial Econometrics III

Winter/Spring 2019
Overview
 The key point

 One open puzzle

 Cointegration approaches to forecasting interest rates

 Yield-curve based factor models

 Principal component modelling

 Excess bond return predictability

 Dynamic linear affine term structure models in forecasting

 The role of macroeconomic information

 A puzzle, predictability from past forward rates: the «tent»
Lecture 3: Forecasting interest rates – Prof. Guidolin 2
The key point
 The time-t term structure contains substantial information
about changes in the slope and curvature
 However, level is close to a random walk process
 Forecasts that imply substantial predictable variation in
expected excess returns may point to overfitting
 Forecasting future, riskless interest rates is of obvious importance
to traders, bond portfolio managers, and policy makers
 Adopt asset pricing approach: interest rates are functions of bond
prices, their dynamics can be studied using tools of asset pricing
o The theory is particularly powerful when applied to Treasury yields,
since the underlying assets have fixed payoffs (unlike stocks)
 Using asset pricing a number of key insights on interest rate
forecasting emerge
 Gaussian dynamic term structure models (TSM) are the tool to
describe joint forecasts of future yields, returns, and risk premia
 These models impose no-arbitrage restrictions
Lecture 3: Forecasting interest rates – Prof. Guidolin 3
One (possibly, open) puzzle
 Standard economic explanations of risk premia fail to explain
the behavior of expected excess returns to bonds
 Unfortunately, standard economic explanations of risk premia fail
to explain the behavior of expected excess returns to bonds
o In the data, mean excess returns to long-term Treasuries are positive
o Yet traditional measures of risk exposure imply that Treasury bonds
are not assets that demand a risk premium
o Point estimates of their consumption betas are negative and point
estimates of their CAPM betas are approximately zero
o Although expected excess returns to bonds vary over time, these
variations are unrelated to interest rate volatility or straightforward
measures of economic growth
 The figure displays zero-coupon bond yields btw. 3m and 10Y
o Academics use zero-coupon yields interpolated from Treasuries
o The interpolation is inherently noisy: Bekaert, Hodrick, and Marshall
(1997, JFE) estimate that the std. deviation of measurement error is in
the range of 7 to 9 bps of annualized yield for maturities > 1 year
Lecture 3: Forecasting interest rates – Prof. Guidolin 4
A first peek at the data

Lecture 3: Forecasting interest rates – Prof. Guidolin 5

Cointegration approaches
 Standard tests reveal that riskless yields are (near) unit-root,
possibly cointegrated but that spreads are stationary
 A glance at the figure suggests that yields are cointegrated: spreads
between yields on bonds of different maturities are mean-
reverting, but the overall level of yields is highly persistent
 Therefore term spreads tend to be stationary
 A robust conclusion of the literature is that standard tests cannot
reject the hypothesis of a unit root in any of these yields
o In an economic perspective it is easier to assume that yields are
stationary and highly persistent rather than truly nonstationary
o Unit root interest rates carry a positive prob. of falling below -100%
 Campbell and Shiller (1987, JPE) motivate a cointegration approach
 They make the simplifying assumption that the weak form of the
expectations hypothesis maturity-
holds so that dependent
constant

Lecture 3: Forecasting interest rates – Prof. Guidolin 6

Cointegration approaches
 If investors have information not captured in the history of
short rates, then spreads forecast changes in short rates
 The spread between the yield on an n-period bond and a 1-period
yield is then

 Spreads are I(0) if one-period yields are I(1).

 Campbell and Shiller examine monthly observations of one-month
and 20-year bond yields over the period 1959 to 1983
 They cannot reject the hypotheses that yields are I(1) and spreads
are I(0) and hence advocate an error-correction model (ECM):

o Investors impound their information about future short-term rates in

the prices of long-term bonds
o If investors have information about future rates not captured in the
history of short rates, then spreads forecast changes in short rates
Lecture 3: Forecasting interest rates – Prof. Guidolin 7
Yield-curve based factor models
 The ECM approach does not recognize that period-t bond yields are
determined based on all information that investors have about
future interest rates, possibly beyond short-term rates
 This fact leads to a more parsimonious approach to modeling
interest rates than an ECM
 Assume that all of the information that determines investors’
forecasts at t can be summarized by a p-dimensional state vector xt

 The yield on the LHS cannot be a function of anything other than

the state vector, since only xt shows up on the RHS,
 Stack time-t yields on bonds with different maturities in a vector yt
 Assume there exists an inverse function s.t.
 The inverse function exists if yields contain the same information
as xt  the rank of ∂f/∂xt must be p
Lecture 3: Forecasting interest rates – Prof. Guidolin 8
Yield-curve based factor models
 If each element of xt has its own unique effect on the time-t
yield curve, the yield curve can be inverted to infer xt
o A necessary condition is that there are at least p yields in the vector yt
o There are technical conditions associated with this result, but the
intuition is that if each element of xt has its own unique effect on the
time-t yield curve, the yield curve can be inverted to infer xt
o Put differently, the time-t yield curve contains all information
necessary to predict future values of xt
 This allows us to write that is determined by
the mappings from factors to expected future one-period yields and
excess bond returns
 A possible interpretation is that both xt and yt must follow first-
order Markov processes
 xt is Markov because it is defined as the set of information relevant
to forming conditional expectations and if information at time t
other than xt were helpful, then investors could use it
Lecture 3: Forecasting interest rates – Prof. Guidolin 9
Yield-curve based factor models
 When yields follow a Markov process, the time-t dynamics
(e.g., regimes) can be backed out of time-t yields
o When forming forecasts as of time-t, the use of information other than
time-t yields requires a compelling justification
o Such information includes yields dated prior to t, measures of
inflation, central bank policy announcements, and economic activity
o All of information in these variables is embedded in time-t yield curve
o However, if there is a good reason to believe that the mapping from
current yields to expected future yields is unstable over time, while
the mapping from, say, current inflation to expected future yields is
stable, then it makes sense to use other data to forecast yields
 It is also important to recognize that a Markov process for yields
does not imply that yields follow a first-order VAR
 Markov processes may be nonlinear, e.g., there may be regime shifts
 Even if a VAR(1) is a reasonable model of yields, we must decide
how to compress the information in the cross-section of yields
Lecture 3: Forecasting interest rates – Prof. Guidolin 10
Principal component forecasting
 Yields are commonly summarized using principal components
o Since yields on bonds of different maturities are highly correlated, it
does not make sense to estimate unrelated regressions for each yield
 A standard approach is to extract common factors from yields and
apply a VAR to the factors
o Forecasts of individual yields are determined by the cross-sectional
mapping from the vector to the individual yield
 Following Litterman and Scheinkman (1991, JFI), researchers often
summarize term structures by a small set of linear combinations
 The first few principal components of the covariance matrix of
yields capture almost all of the variation in the term structure
 Std deviations of residuals from using three principal components
range from 5 to 11 bps which is roughly the same range as the
measurement error described by Bekaert et al. (1997, JFE)
 These first three principal components are commonly called “level,”
“slope,” and “curvature”, respectively
Lecture 3: Forecasting interest rates – Prof. Guidolin 11
Principal component forecasting
 Yields are commonly summarized using principal components

Lecture 3: Forecasting interest rates – Prof. Guidolin 12

Principal component forecasting
 Yields are commonly summarized using principal components

Lecture 3: Forecasting interest rates – Prof. Guidolin 13

Principal component forecasting
 The level factor tends to be unpredictable
 A VAR(1) applied to PCs allows us to use time-t principal
components to predict time-(t+k) principal components

 Other models, for instance MSVAR, are possible or even worthy

 The mapping
from PCs to yields
translates these
forecasts to
expected future
yields
 No statistical
evidence that
changes in level
are forecastable
Lecture 3: Forecasting interest rates – Prof. Guidolin 14
Principal component forecasting
 Slope and curvature are strongly forecastable
 At a quarterly horizon, about 20 (30) percent of the variation in
slope (curvature) is predictable
 PCs other than the first three do not contribute much to
forecasts of slope and curvature

Lecture 3: Forecasting interest rates – Prof. Guidolin 15

Principal component forecasting
 Forecasts of future yields using current yields are necessarily
also forecasts of expected log returns to bonds
 Duffee (2011) takes this approach and concludes that the model
works well in pseudo out-of-sample forecasting
 Diebold and Li (2006, JoE) build a dynamic version of the term
structure introduced by Nelson and Siegel (1987, JBus)
o The cross-section of the term structure is summarized by the level,
slope, and curvature factors of the Nelson-Siegel model
o These three factors are assumed to follow a VAR(1)
 Diebold and Li find that the dynamics with the best pseudo out-of-
sample properties are those in which level, slope, and curvature
follow univariate AR(1) processes
 A popular forecasting approach uses a VAR that includes both
compressed information from the term structure and compressed
information from a large panel of macro variables
 Should we forecast future yields or future bond returns?
Lecture 3: Forecasting interest rates – Prof. Guidolin 16
Excess bond return predictability
 If changes in long-term rates are unpredictable, then long bond
excess returns must be predictable and related to term premia
o A derivation in Campbell and Shiller (1987, JPE) shows that there is
no difference

o For reasonably long maturities, variations over time in the LHS are
very close to variations in the 1st PC, hence close to unforecastable
o Therefore the RHS must also be unforecastable with time-t yields
o The first term on the right side is a measure of the slope of the term
structure, which varies widely over time
o Because the sum on the RHS is unforecastable, the second term,
excess returns to the bond, must also be strongly forecastable and
positively correlated with the slope of the term structure
 This implication is confirmed with excess return regressions from
CRSP constructed by subtracting the return to the shortest-
maturity portfolio

Lecture 3: Forecasting interest rates – Prof. Guidolin 17

Excess bond return predictability
 The level of the term structure is unrelated to future excess returns
 A less-steep slope (larger value of the second principal component)
corresponds to lower future excess returns
 Less clear are the links
btw. other PCs and future
excess returns
o There is strong statistical
evidence that greater
curvature predicts lower
excess returns
o At the monthly horizon,
5% of the variation is
predictable, 10% at
the quarterly horizon

Lecture 3: Forecasting interest rates – Prof. Guidolin 18

Dynamic, affine term structure models
 For tractability, assume that interest rates are linear and homoske-
dastic with Gaussian shocks and rule out arbitrage opportunities
o Why a linear, homoskedastic model?
o It is easy to find evidence of nonlinear, non-Gaussian dynamics
o Gray (1996, JFE) concludes that a model of time-varying mean
reversion and time-varying GARCH effects fits the dynamics of the
short-term interest rate
o A zero bound imposes nonlinear dynamics on yields that bind
 Although they have fixed conditional variances, this is typically not
a concern of the forecasting literature that focuses on predicting
yields and excess returns rather than conditional second moments
 The short rate, rt, is a function of p state variables, xt,
 The state vector has first-order VAR(1) Markov dynamics

 Given a unique SDF, no-arbitrage implies

Lecture 3: Forecasting interest rates – Prof. Guidolin 19
Dynamic, affine term structure models
 In a linear affine Gaussian homoscedastic model, the state
follows a VAR(1) process and risk prices are linear in the state
 The SDF is assumed to have the log-linear form

o t is a vector with time-varying prices of risk, a function of the state,

o Bonds are priced using the equivalent martingale dynamics:

 At this point, bond prices can be written as

(difference equation)

o Zero-coupon yields are written as

o Excess return:
Lecture 3: Forecasting interest rates – Prof. Guidolin 20
Identification of linear affine TSM
 Restrictions and/or normalizations need to be imposed to
identify a linear affine dynamic TSM
 As far estimation is concerned, see lecture notes from a.a. 2015 and
earlier posted on the course web site
 One important problem is left: the state vector is arbitrary, in the
sense that an observationally equivalent model is produced by
scaling, rotating, and translating the state vector
o Define such a transformation as

o Dai and Singleton (2000, JF) show that an observationally equivalent

model replaces xt with x*t , and replaces the parameters with

 Many ways to identify the state vector and thus the parameters
 One way is to restrict the K matrix to a diagonal matrix, set μ to
zero, and set the diagonal elements of Σ equal to one
 These define a rotation, translation, and scaling, respectively
Lecture 3: Forecasting interest rates – Prof. Guidolin 21
Identification of linear affine TSM
 Two common identification restrictions are based on either
some yields being true or on principal components
 Other approach equates xt with linear combinations of yields:
consider any p × d matrix L with rank p, d = number of maturities

 As long as ℒB is invertible, this defines a new state vector

o A simple example is a diagonal ℒ with p diagonal elements equal to
one and the remainder equal to zero
o This choice produces a state vector equal to p “true” yields, or yields
uncontaminated by measurement error
 Other choice of ℒ is based on PCs of yields:
 Then the factors correspond to “level,” “slope,” and so on
 Regardless of the choice of ℒ, this kind of state vector rotation
emphasizes that a TSM is really a model of one set of derivative
instruments (yields) explaining another set of derivative
instruments (more yields)
Lecture 3: Forecasting interest rates – Prof. Guidolin 22
Do no arbitrage restrictions help?
 General consensus is that while restrictions on risk premia
estimates helps forecasting, no arbitrage restrictions do not
 It seems natural to estimate unrestricted risk premia, λ0 and λ1, but
Joslin, Singleton, and Zhu (2011, RFS) and Duffee (2011) show that
this version of the no-arbitrage model is of no value in forecasting
o Joslin et al.: no-arb, in the absence of restrictions on risk premia, have
no bite when estimating conditional expectations of the state vector
o These are determined by μ and K of the true, or physical measure
o No-arb restrictions boil down to existence of equivalent martingale
dynamics, and when risk premia are unrestricted, the parameters μq
and Kq are unrelated to their physical-measure counterparts
o The two measures share volatility parameters, in a Gaussian setting
these parameters do not affect ML estimates of μ and K.
o No-arb restrictions affect the mapping from the state to yields: A, B
o Duffee argues these parameters can be estimated with very high
precision even if no-arb restrictions are not imposed, since the
measurement equation amounts to a regression of yields on other
yields Lecture 3: Forecasting interest rates – Prof. Guidolin 23
Do no arbitrage restrictions help?
 Restrictions on risk premia increase the precision of estimates of
physical dynamics
 The reason is that equivalent-martingale dynamics are estimated
with high precision, and risk premia restrictions tighten the
relation between physical and equivalent-martingale dynamics
 E.g., given the task of estimating λ0 and λ1, one is to set to zero any
parameters that are statistically indistinguishable from zero
o Other approaches rely on information criteria, Bayesian shrinkage,
weighted averages btw. physical and risk-neutral values of zero
o Christensen, Diebold, and Rudebusch (2011, JoE) propose a dynamic
model of the three factors of Nelson and Siegel, subject to no arbitrage
 However, the main avenue seems clear: to bring macroeconomic
information into the problem to derive the SDF from fundamentals
 Macro-finance models follow Ang and Piazzesi (2003, JME) by
expanding the measurement equation of the yields-only framework
 Assume we observe some variables at time t, stacked in a vector ft
Lecture 3: Forecasting interest rates – Prof. Guidolin 24
Macroeconomic factors in no-arbitrage models
 Usually it contains macro variables, but may contain survey data:

o No-arb restrictions apply only to the vector Ay and the matrix By

o The key assumption is that the same state vector that determine the
cross-section of yields also determine the additional variables
o A difficulty with macro-finance models is that the macro variables of
interest, such as inflation, aggregate consumption, and output growth,
are not spanned by the term structure
o Aside from measurement error, regressions of the macro factors on p
linear combinations of contemporaneous yields should produce R2s of
1!

Lecture 3: Forecasting interest rates – Prof. Guidolin 25

Macroeconomic factors in no-arbitrage models
o The hypothesis that the term structure contains all information
relevant for forecasting future macro vars is overwhelmingly rejected
o The term structure has less information about future inflation and IP
growth than is contained in the single lag of the macro variable
 Duffee (2011, RFS) and Joslin, Priebsch, and Singleton (2014, JF)
develop a restricted Gaussian no-arbitrage models in which no
linear combinations of yields can serve as the model’s state vector
 They show that hidden factors must exist
 The term structure of bond yields is not a first-order Markov
process, i.e., investors have information about future yields and
future excess returns that is not impounded into current yields
 Hidden factors can explain the low R2s as long as the factor(s) that
are hidden are revealed in macroeconomic data
o E.g., imagine that economic growth suddenly stalls
o Investors anticipate that short-term rates will decline
o But investors’ risk premia also rise because of the downturn
Lecture 3: Forecasting interest rates – Prof. Guidolin 26
Macroeconomic factors in no-arbitrage models
 Unfortunately, the causal linkages btw. macroeconomic
fundamentals and excess bond returns are modest at best
o These effects move long-term bond yields in opposite directions and if
they happen to equal each other in magnitude, the current term
structure is unaffected
o Nothing in the term structure
predicts what happens next
to either economic growth
or bond yields
o However, lower growth should
predict higher excess returns
 At quarterly frequency, excess
Treasury bond returns are
countercyclical
o On average, the nominal yield
curve slopes up and hence
expected excess returns to
Treasury bonds are positive
Lecture 3: Forecasting interest rates – Prof. Guidolin 27
The forecasting power of forward rates: a puzzle
 Oddly, past forward rates contain information on risk premia
o Estimated correlations in the table imply that they should be negative
 Is there really information about future excess returns that is not
captured by the current term structure?
 Cochrane and Piazzesi (2005, AER) find that 5-month forward rates
constructed with yields on maturities of 1 through 5 years contain
substantial information about excess returns over the next year

 They conclude that lags of forward rates contain information about

the excess returns that is not in month t forward rates.
 Even if there are hidden factors, it is hard to understand why the
information is hidden from the time-t term structure but not
hidden in earlier term structures
 CP suggest measurement error in yields that is averaged away by
including additional lags of yields
Lecture 3: Forecasting interest rates – Prof. Guidolin 28
The forecasting power of forward rates: a puzzle

 For the 1964 through 2003 sample, including lagged forward rates raises
R2 from 5.5 to 11%; the null that the coefficients on the forward rates are
all zero is overwhelmingly rejected
 However, over the full sample, the incremental explanatory power of the
forward rates is more modest
Lecture 3: Forecasting interest rates – Prof. Guidolin 29
Conclusion
 Finance theory provides some guidance when forming forecasts of
future interest rates
 The Holy Grail of this literature is a dynamic model that is
parsimonious owing to economically-motivated restrictions
 The requirement of no-arbitrage is motivated by economics, but by
itself it is too weak to matter
 Economic restrictions with bite require, either directly or indirectly,
that risk premia dynamics be tied down by economic principles
 No restrictions from workhorse models of asset pricing appear to
be consistent with the observed behavior of Treasury bond yields
 Another open question is whether any variables contain
information about future interest rates that is not already in the
current term structure
o Recent empirical work suggests that both lagged bond yields and
certain macroeconomic variables reflect incremental information, but
the robustness of these results is not yet known
Lecture 3: Forecasting interest rates – Prof. Guidolin 30