Term Premia and Interest Rate Forecasts in Affine Models
by
Gregory R. Duffee
Citation
Title:
Term Premia and Interest Rate Forecasts in Affine Models
Author:
Gregory R. Duffee
Year:
2002
Publication:
The Journal of Finance
Volume:
57
Issue:
1
Start Page:
405
End Page:
443
Publisher:
Language:
English
URL:
Select license:
Select License
DOI:
PMID:
ISSN:
Updated: February 14th, 2013
Abstract:
Term Premia and Interest Rate Forecasts in Affine Models
GREGORY R. DUFFEE"
ABSTRACT
The standard class of affine models produces poor forecasts of future Treasury yields. Better forecasts are generated by assuming that yields follow random walks. The failure of these models is driven by one of their key features: Compensation for risk is a multiple of the variance of the risk. Thus risk compensation cannot vary independently of interest rate volatility. I also describe a broader class of models. These "essentially affine" models retain the tractability of standard models, but allow compensation for interest rate risk to vary independently of interest rate volatility. This additional flexibility proves useful in forecasting future yields.
CANWE USE FINANCE THEORY to tell us something about the empirical behavior of Treasury yields that we do not already know? In particular, can we sharpen our ability to predict future yields? A longestablished fact about Treasury yields is that the current term structure contains information about future term structures. For example, longmaturity bond yields tend to fall over time when the slope of the yield curve is steeper than usual. These predictive relations are based exclusively on the timeseries behavior of yields. To rule out arbitrage, the crosssectional and timeseries characteristics of the term structure are linked in an internally consistent way. In principle, imposing these restrictions should allow us to exploit more of the information in the current term structure, and thus improve forecasts. But in practice, existing noarbitrage models impose other restrictions for the sake of tractability; thus their value as forecasting tools is a priori unclear.
I examine the forecasting ability of the affine class of term structure models. By "affine," I refer to models where zerocoupon bond yields, their physical dynamics, and their equivalent martingale dynamics are all affine functions of an underlying state vector. A variety of nonaffine models have been developed, but the tractability and apparent richness of the affine class has led the finance profession to focus most of its attention on such models.
Although forecasting future yields is important in its own right, a model that is consistent with finance theory and produces accurate forecasts can make a deeper contribution to finance. It should allow us to address a key
* Haas School of Business, University of California. I thank Mark Fisher for many helpful conversations and Jonathan Berk; Rob Bliss; Qiang Dai; Darrell Duffie; an anonymous referee; and seminar participants at Berkeley, Yale, Stanford, and an NBER Asset Pricing meeting for useful comments.
issue: explaining the time variation in expected returns to assets. In the context of the term structure, explaining timevarying returns means explaining the failure of the expectations hypothesis of interest rates. Put dif ferently, we would like to have an intuitive explanation for the positive correlation between the yield curve slope and expected excess returns to long bonds. If a model produces poor forecasts of future yields (and thus poor forecasts of future bond prices), it is unlikely that the model can shed light on the economics underlying the failure of the expectations hypothesis.
The first main conclusion reached here is that the class of affine models studied most extensively to date fails at forecasting. I refer to this class, which includes multifactor generalizations of both Vasicek (1977) and Cox, Ingersoll, and Ross (1985), and is analyzed in Dai and Singleton (2000), as "completely affine." I fit general threefactor completely affine models to the Treasury term structure over the period 1952 through 1994. Yield forecasts produced with these estimated models are typically worse than forecasts produced by simply assuming yields follow random walks. This conclusion holds for both insample forecasts and outofsample (1995 through 1998) forecasts.
Even more damning is the way in which the estimated models fail. They produce yield forecast errors that are strongly negatively correlated with the slope of the yield curve. In other words, the models fail to replicate the key empirical relation between expected returns and the slope of the yield curve; their underestimation of expected excess returns to long bonds tends to be largest when the slope of the term structure is steep.
This failure is a consequence of two features of the Treasury term struc ture, combined with a restriction built into completely affine models. The first feature is that the distribution of yields is not strongly skewed; yields vary widely over time around both sides of their sample means. The second feature is that, while the average excess return to Treasury bonds is not much greater than zero, the slope of the term structure predicts a relatively large amount of variation in excess returns to bonds. Fama and French (1993) note that this implies the sign of predicted excess returns to Treasury bonds changes over time.
Completely affine models do not simultaneously reproduce these two fea tures of term structure behavior. The key restriction in these models is that compensation for risk is a fixed multiple of the variance of the risk. This structure ensures that the models satisfy a requirement of noarbitrage: Risk compensation goes to zero as risk goes to zero. But because variances are nonnegative, this structure restricts the variability of the compensation that investors expect to receive for facing a given risk. The compensation is bounded by zero; therefore it cannot switch sign over time.
As will be made clear in the paper, the only way this framework can pro duce expected excess returns with low means and high volatilities is for some underlying factors driving the term structure to be highly positively skewed. But this strong positive skewness is inconsistent with the actual distribution of yields. Thus completely affine models can fit either of these features of Treasury yields, but not both simultaneously.
All is not lost, however. The second main conclusion of this paper is that the completely affine class can be extended to break the link between risk compensation and interest rate volatility. This extension from the com pletely affine class to the "essentially affine" class described here is costless, in the sense that the affine timeseries and crosssectional properties of bond prices are preserved in essentially affine models. The existence of extensions to the completely affine class is not new (Chacko (1997) constructs a general equilibrium example), but this paper is the first to describe and empirically investigate a general, very tractable extension to completely affine term struc ture models. I find that essentially affine models produce more accurate yield forecasts than completely affine models, both in sample and out of sample. However, there is a tradeoff between flexibility in forecasting fu ture yields and flexibility in fitting interest rate volatility.
The paper is organized as follows. The structure of affine models is dis cussed in detail in Section I. Section I1 explains intuitively why completely affine models work poorly. Section I11 describes the estimation technique. Section IV presents empirical results and Section V concludes.
I. Affine Models of the Term Structure
A. Affine Bond Pricing
The core of affine term structure models is the framework of Duffie and Kan (1996). Their model, summarized here, describes the evolution of bond prices under the equivalent martingale measure. Uncertainty is generated by n Brownian motions, w, = (w,,,,..., G,,,)'. There are n state variables, denoted X, = (X,,,, . . . ,X,,,,)'. The instantaneous nominal interest rate r, is affine in these state variables:
where 6, is a scalar and 6 is an nvector. The evolution of the state variables under the equivalent martingale measure is
where K~ and C are n x n matrices and (~0)~
is an nvector. The Q superscript distinguishes parameters under the equivalent martingale measure from corresponding parameters under the physical measure. The matrix S, is diagonal with elements
where pi is an nvector and ai is scalar. It is convenient to stack the pi vectors into the matrix p, where pi is row i of p. The scalars cui are stacked in the nvector a. The following discussion assumes that the dynamics of (1) are well defined, which requires that cri + PIX, is nonnegative for all i and all possible X,. Parameter restrictions that ensure these requirements are in Dai and Singleton (2000).
Denote the timet price of a zerocoupon bond maturing at time t + r as P(X,,r). Duffie and Kan (1996) show
where A(r) is a scalar function and B(r) is an nvalued function. Thus, the bond's yield is affine in the state vector:
The functions A(T) and B(r) can be calculated numerically by solving a series of ordinary differential equations (ODEs).
B. The Price of Risk and Expected Returns to Bonds
The model is completed by specifying the dynamics of X, under the phys ical measure, which is equivalent to specifying the dynamics of the price of risk. Denote the state price deflator by T,. The relative dynamics of T, are
where W, follows a Brownian motion under the physical measure. Element i of the nvector A, represents the price of risk associated with W,,i. The dynamics of X, under the physical measure can be written in terms of A, and (1):
Instantaneous bondprice dynamics can be written as
dP(X,,7) = (r, + e,,,)dt + v,,,dW,, P(Xt,71
where e,,, denotes the instantaneous expected excess return to holding the bond. An application of Ito's lemma combined with the structure of the ODEs in Duffie and Kan (1996) reveals
Equation (7) says that variations over time in expected excess returns are driven by variations in both the volatility matrix S, and the price of risk vector A,. A parametric model for bondyield dynamics requires a functional form for A,. This form should be sufficiently flexible to capture the empir ically observed behavior of expected excess returns. Thus, to motivate the choice of functional form for A,, we briefly review evidence on the behavior of bond returns.
A large literature documents that expected excess returns to Treasury bonds (over returns to shortterm Treasury bills) are, on average, near zero, and vary systematically with the term structure.l When the slope of the term structure is steeper than usual, expected excess returns to bonds are high, while expected excess returns are lowoften negativewhen the slope is less steep. Thus the ratio of mean expected excess bond returns to the stan dard deviation of expected excess bond returns is low.
Earlier work has also shown that the shape of the term structure is re lated to the volatility of yields.2 However, the slopeexpected return relation is not simply proxying for a volatilityexpected return relation. Supporting evidence is in Table I, which reports results from regressions of excess monthly bond returns on the lagged slope of the term structure and lagged yield volatility. Monthly returns to portfolios of Treasury bonds are from the Cen ter for Research in Security Prices. Excess returns to these portfolios are produced by subtracting the contemporaneous return to a threemonth Trea sury bill. The slope of the term structure is measured by the difference be tween monthend fiveyear and threemonth zerocoupon yields. The yields are interpolated from coupon bonds using the technique of McCulloch and Kwon (19931, as implemented by Bliss (1997).3 Yield volatility is the stan dard deviation of the fiveyear zerocoupon yield, measured by the square root of the sum of squared daily changes in the yield during the month.
The sample period is July 1961 through December 1998. The results in Table I reveal that month t's volatility has no statistically significant pre dictive power for excess bond returns in month t + 1.By contrast, all of the estimated slope parameters are significant at the 10 percent level and half are significant at the 5 percent level. In addition, the variation in predicted excess returns is large relative to mean excess returns. Consider, for exam ple, bonds with maturities between three and four years. The mean excess return is 7 basis points per month, while the standard deviation of predict able excess returns is roughly 25 basis points. Armed with this information about the empirical behavior of bond returns, we now discuss three different parameterizations of A,.
The literature is too large to cite in full here. Early research includes Fama and Bliss (1987). Two standard references are Fama and French (1989, 1993).
This literature is also too large to cite in full. Chan et al. (1992) examine the sensitivity of volatility to the level of shortterm interest rates. Andersen and Lund (1997) refine their work by decomposing the variation in interestrate volatility into a component related to the level of shortterm interest rates and a stochastic volatility component.
I thank Rob Bliss for providing me with the yield data.
Table I
Regressions of Excess Returns to Treasury Bonds,
July 1961 through December 1998
Monthly excess returns to portfolios of Treasury bonds are regressed on the previous month's termstructure slope and an estimate of the interest rate volatility during the previous month. The slope of the term structure is measured by the difference between fiveyear and three month zerocoupon yields (interpolated from coupon bonds). Monthly volatility is measured by the square root of the sum of squared daily changes in the fiveyear zerocoupon bond yield. Asymptotic tstatistics, adjusted for generalized heteroskedasticity, are in parentheses. There are 449 monthly obsen~ations.
Coefficient on Maturity Mean Excess Return .Std. Dev. of Fitted
(years) Slope Volatility Excess Rets
C. Completely Affine Models
Fisher and Gilles (1996) and Dai and Singleton (2000) adopt the follow ing form for A,. Let h, be an nvector. Then the price of risk vector A, is given by
This class nests multifactor versions of Vasicek (1977) and Cox, Ingersoll, and Ross (1985; hereafter CIR). The main reason for the popularity of this form is that the vector S,h, is affine in X,. This implies affine dynamics for X, under both the equivalent martingale and physical measures. Affine dy namics of X,under the physical measure allow for closedform calculation of various properties of conditional densities of discretely sampled yields. These properties are discussed in detail in Duffie, Pan, and Singleton (2000) and Singleton (2001). Of less importance is the fact that A',&, which is the in stantaneous variance of the log state price deflator, is also affine in X,. This latter property motivates the term "completely affine," as discussed in the next subsection.
This structure imposes two related limitations on A,. First, variation in the price of risk vector is completely determined by the variation in St. There fore, variations in expected excess returns to bonds are driven exclusively by the volatility of yields, an implication that appears inconsistent with the evidence in Table I. Second, the sign of element i of A, is the same as that of element i of A,, because the diagonal elements of S, are restricted to be nonnegative. The importance of this limitation will be clear in Section 11.
D. Essentially Affine Models
The essentially affine class nests the completely affine class. We first de fine the elements of a diagonal matrix Stpas
s..= (ai + P;X,)~'~,if inf(ai + PIXt) > 0;
~(LL)
otherwise.
Thus, if diagonal element i of S, is bounded away from zero, its reciprocal is diagonal element i of S;. For any diagonal element of S, with a lower bound of zero (whether or not it is accessible given the dynamics of X,), the associated element of SF is set to zero. Therefore the elements of S, do not explode as the corresponding elements of S, approach zero.
The form of A, used in the essentially affine model is
where A, is an n x n matrix. This form shares with (9) two important prop erties. First, if approaches zero, A, does not go to infinity. Second, S,A, is affine in X,. Therefore, the physical dynamics of X, are affine.
There are three important differences between (9) and (10). First, with A, # 0, &A, is not affine in X,. Therefore, this model is not completely affine, but the variance of the state price deflator does not affect bond prices. This motivates the term "essentially affine." Second, the tight link between the price of risk vector and the volatility matrix is broken. The essentially affine setup allows for independent variation in prices of risk, which is the kind of flexibility needed to fit the empirical behavior of expected excess returns to bonds. Third, the sign restriction on the individual elements of A, is removed.
For future reference, we require an expression for the physical dynamics of X,. Substitute (10) into (6) and define Ias the n X n diagonal matrix with Ii; = 1if SGii)# 0, Ii; = 0 if St(ii) = 0. Then the physical dynamics in the essentially affine model are
The Journal of Finance Combining terms and denoting element i of A, by Ali, (11)can be written as dX, = (KO KX,)dt + CS,d W,, (124 where
and
E. An Essentially Affine Example
The following twofactor model illustrates the essentially affine frame work. The instantaneous interest rate r, follows a Gaussian process and there is some factor f, that follows a squareroot process. It is convenient to begin by modeling their dynamics under the physical measure. Under this measure, the processes are independent:
The model is closed with a description of the dynamics of the market price of risk. If we adopt the completely affine version in (9), the result is the Vasicek (1977) model for r,. In such a setup, the variable ft is irrelevant for bond pricing, and we are left with a standard onefactor Gaussian model.
If, however, we use the essentially affine specification for the market price of risk, the factor ft can affect bond prices, even though it cannot affect the path of r,. The reason is that the compensation that investors require to face the risk of W,,, can vary with f,. The essentially affine model specifies the price of risk A, as
The dynamics of the state price deflator are therefore
The dynamics of r, and f, under the equivalent martingale measure are, from (12a) and (12b),
where fQand r are the means of f, and r, under the equivalent martingale measure.
There are three important differences between this description of bond price dynamics and the standard Vasicek model. First, the instantaneous interest rate r, affects the price of interest rate risk, through the parameter A,!,,,. In Vasicek, the price of interest rate risk is constant. Second, there is a source of uncertainty in bond prices that is independent of the physical dynamics of r,. The factor f, affects bond prices through the parameter A,(,,). Chacko (1997) builds an affine term structure model expressly designed to exhibit this second feature, and my example was inspired by his (substan tially more complicated) model. We will see in Section IV that this kind of feature is critical to understanding the actual dynamics of Treasury bond yields. Third, the price of risk associated with innovations in W,,, can change sign, depending on the level of the factor ft.
The essentially affine structure of A,, although more flexible than the completely affine structure, nonetheless imposes limits on the possible dy namics of bond prices. Note that one element of K, which is the first matrix on the righthand side of (13)) is the same as the corresponding element of K~,
which is the first matrix on the righthand side of (14). Element (1,2) must be zero under both the physical and equivalent martingale measures. Otherwise, the drift off, at f, = 0 could be negative (because it would depend on r,), which cannot be allowed because 6 enters into St.
To free up this element, and thus allow for a more flexible specification of the price of risk, we could model f, as a Gaussian process. An example of such a model is Fisher (1998). By contrast, if both f, and r, were modeled as squareroot diffusion processes, the essentially affine structure of A would be identical to the completely affine structure. This illustrates a more gen eral point noted by Duffie and Kan (1996) and Dai and Singleton (2000)) and that we will see in Section IV. With the affine setup, there is a tradeoff between constructing a model that can capture complicated dynamics in vol atilities and a model that can capture complicated dynamics in expected returns.
F. Semiaffine Models
Duarte (2000) chooses an alternative generalization of completely affine models. Let A, be an nvector. The price of risk vector is
With this form, elements of A, can switch sign over time, but they cannot move independently of St. As noted in Section I.C, this latter feature ap pears inconsistent with the empirical evidence. Thus, at first glance, it seems that the semiaffine setup allows for some, but not all, of the flexibility of the essentially affine setup. However, there are parameterizations of St for which the semiaffine model offers more flexibility than does the essentially affine model. One example is the multifactor CIR model, which is the focus of Duarte's empirical work. The semiaffine form of A, implies nonaffine dynamics of X, under the physical measure. Therefore, as noted by Duarte (2000)) approximation or simulation techniques are typically necessary to reproduce the properties of discretely sampled yields.
G. A Canonical Form for Essentially Affine Models
There are a variety of normalizations that can be imposed on affine mod els. I follow the lead of Dai and Singleton's (2000) canonical completely af fine model. They normalize C to the identity matrix. The form of the model is determined by the total number of state variables n and the number of state variables that affect the instantaneous variance of X,, denoted m. The state vector is ordered so these are the first m elements ofX,. The resulting model is an A,(n) model. They set the first m elements of a to zero and the remaining n m elements to one. Their version of (2) is
where for i = m + 1,.. . , n,
Using their framework, we can write the diagonal elements of SF and Ias
Note that in (ll),the matrix A, shows up only in the term IA,X,. Therefore, we can normalize the first m rows of A, to zero. Now reconsider (7)) the instantaneous expected excess return to holding a bond with re maining maturity T. From (10)) (15)) and (16)) in the canonical form, this can be written as
In (171, 0, is an mvector of zeros. The O,,, matrices are defined similarly. The submatrix Ma is a diagonal matrix with the ith diagonal element equal to element i of A,. Row i of Mbis given by the first m elements of the vector A,(,+,,P&+,. The submatrix L consists of rows (m + 1)through n of A,.
The additional flexibility of the essentially affine model in fitting time variation in expected excess returns to bonds is captured by the matrix L. In a completely affine setup, L is a zero matrix. Therefore, any elements of X, that do not affect the instantaneous volatility ofX, (i.e., elements m + 1,... ,n) also cannot affect instantaneous expected excess returns to bonds. When L is nonzero, any such elements of X, can affect expected excess returns. In addition, L provides a mechanism for all other elements in X, to affect ex pected returns through a channel other than Ma or Mb.
If all elements of X, affect the instantaneous volatility (i.e., a correlated multifactor CIR model, or what Dai and Singleton (2000) call an A,(N) model), there is no L matrix. Therefore, the essentially affine model generalizes the completely affine model only when there is at least one element in X, that does not affect the instantaneous volatility of X,.
11. The Intuition Behind the Failure of Completely Affine Models
A successful model of the term structure should be consistent with the variety of termstructure shapes observed in the data. In other words, given the model's parameters, for each observed shape, there should be a valid state vector X, that can generate it. In addition, the model should reproduce the empirically observed patterns in expected returns to bonds, or equiva lently, produce forecasts of future yields that subsume the forecasting infor mation in the slope of the term structure. This section explains that completely affine models fit to the historical behavior of Treasury yields will not simul taneously achieve both of these goals.
For our purposes, the key features of the excess returns to bonds are that they are, on average, small, and exhibit substantial predictable variation. Recall from Section I that e,,, denotes the instantaneous expected excess return to a bond with maturity 7.Although we do not observe instantaneous returns, the evidence in Table I suggests that the ratio E(e,,,)/.\I is smallwell below onefor all 7.(This ratio is the inverse of the coefficient of variation for e,, ,.)
We will see below that completely affine models can be parameterized to produce low values of E(e,,,)/dm for all T. However, completely affine models can fit this behavior only by giving up the ability to fit a wide range of termstructure shapes. Conversely, they can be parameterized to fit ob served termstructure shapes, but not the behavior of expected excess re
turns. The intuition underlying this result is best seen in two steps. We first examine the behavior of onefactor completely affine models. Then we will see that the important properties of onefactor models carry over to multi
factor models.
A. Onefactor Models
The intuition in a completely affine onefactor model is straightforward. Expected instantaneous excess bond returns, e,,, , are proportional to the factor's variance; hence they are bounded by zero. For a random variable that is bounded by zero to have a standard deviation substantially larger than its mean, it must be highly skewed. This high skewness is a tight restriction on the admissible values of e,,, , and thus a tight restriction on the admissible values of the factor.
To see this clearly, we work through the math. Our first goal is to repro duce the stylized fact that E(e,,,)/d= is small. We restrict our atten tion to a nonGaussian model, because in a completely affine Gaussian model Var(e,,,) = 0. The model is
From (7), the instantaneous expected excess return to a 7maturity bond is
Therefore, the inverse of the e,, ,'s coefficient of variation is
Equation (18) implicitly imposes B(~)ah, < 0, which is the condition that mean excess bond returns are positive. We set E(e,,,)/d= = 0.3, which is a typical ratio for predictable excess returns in Table I. We set the uncon ditional mean and standard deviation of the instantaneous interest rate to
5.5 percent and 2.9 percent, respectively. These values correspond to the mo ments of the threemonth bill yield over 1952 through 1998. In this model, Var(r,) =Var(x,). Plugging the standard deviation into (18) produces 0 = 0.87 per cent. Therefore 6, = 4.63 percent.
The requirement that the mean of x, is small relative to its standard de viation gives the model little flexibility in producing shortterm interest rates that are below average. The instantaneous interest rate r, cannot be less than 6,, or 4.63 percent. But over the period 1952 through 1998, three month yields have ranged from 0.6 to 16 percent. Put differently, the model's parameters and the observed variation in shortterm interest rates over this period imply a range of x, from 4.0 to 11.4; the implied x, is negative in more than 40 percent of the monthly observations. If we parameterized the model to be consistent with the observed distribution of shortterm interest rates (i.e., nonnegative implied x,), we would require 0 > 4.9 percent. But
then the ratio E(e,,,)/d= would exceed 1.6. We can parameterize the model to fit either the expected excess returns or observed term structures, but not both.
We can also think about this model's restriction on the behavior of interest rates in terms of skewness of expected excess returns. In order to produce a small E(e,, ,)/I/=, the model will generate expected excess returns that are always positive, usually very close to zero, and occasionally well above zero. But as noted in Section I, observed expected excess returns are not so posi tively skewed; they range from positive to negative.
B. Multifactor Models
Multifactor models are better at fitting the behavior of expected excess bond returns. For example, it is simple to generate a nearzero value of
E(e,,,)/dm
for a specified maturity, while retaining substantial flex ibility in fitting term structure shapes. All that is required is prices of risk (elements of A,) with different signs. If one element of A, is positive and another negative, then at some maturity 7, the factor loadings will weight these prices of risk such that E(e,, ,) = 0 and Var(e,,,) > 0.
However, completely affine models will not produce nearzero values of ~(e,,~)/dm
for all maturities while also allowing for a wide variety of termstructure shapes. To slightly oversimplify, the intuition is that long maturity bond yields are affected by a single factorthe factor with the greatest persistence under the equivalent martingale measure. Thus, we can use the earlier intuition developed for onefactor models to conclude that multifactor models cannot reproduce the observed behavior of longmaturity yields.
The reason why only a single factor will affect longmaturity bond yields is practical, not theoretical. There are a variety of types of shocks that affect the term structure (e.g., level, slope, twist). Multifactor models capture this variety through factors that die out at different rates under the equivalent martingale measure. In principle, we could construct, say, a twofactor model where both factors affected longbond yields. The only requirement is that the factors have the same speed of mean reversion. But by doing so, we sacrifice the major advantage of multifactor modelsthe ability to fit dif ferent kinds of shocks to the term structure. Thus, such a model would pro duce a poor fit of term structure data relative to a model in which each factor had its own speed of mean reversion.
The failure of completely affine models to fit the empirical behavior of bonds can be seen in the parameter estimates of threefactor completely affine models in Dai and Singleton (2000). They use U.S. dollar interest rate swap yields to estimate the same general threefactor completely affine mod els that are estimated here. I use their data and the parameters of their preferred model to produce implied time series of the state vector and ex pected excess returns to bonds. The results of this exercise, which are not reported in any table, indicate the model captures the combination of low mean and high volatility for expected excess returns. However, in over one quarter of the observations, the implied value of the state vector violates a nonnegativity constraint. The violations tend to occur when the long end of the term structure is well below its average. Thus the results in Dai and Singleton support the conclusion that completely affine models do not simul taneously fit the behavior of expected excess returns to bonds and the ob served term structure shapes.
111. Estimation of Essentially Affine Models
A. Threefactor Affine Models
All of the affine models I estimate have three underlying factors (n = 3). Litterman and Scheinkman (1991) find that three factors explain the vast majority of Treasury bond price movements. This is fortunate, because gen eral threefactor affine models are already computationally difficult to esti mate owing to the number of parameters. Adding another factor would make this investigation impractical. Seven models are estimated: four completely affine models and three essentially affine models. A completely affine model is estimated for each possible number of factors that do not affect the in stantaneous volatility of Xt (from three to zero). Using the canonical form of Section I.G, the estimated models are A,(3) throughA3(3). The other models that are estimated are the essentially affine generalizations of A,(3),A1(3), and A2(3). (Recall A3(3) has no essentially affine generalization.)
The estimated models share the following expressions for the instanta neous interest rate, the physical dynamics ofXt, and the price of risk vector:
Depending on the model, restrictions are placed on the parameters in (19a) through (19d).
B. The Data
I use monthend yields on zerocoupon Treasury bonds, interpolated from coupon bonds using the method in McCulloch and Kwon (1993). Their sam ple, which ends in February 1991, is extended by Bliss (1997). The entire data set covers the period January 1952 through December 1998.4 The range of maturities is three months to ten years.
To perform both insample and outofsample tests, I estimate term structure models using data from 1952 through 1994. The final four years of data are reserved for constructing outofsample forecast errors.
C. The Estimation Technique
I estimate these models using quasi maximum likelihood (QML), which is particularly easy to implement with completely and essentially affine mod els. Although QML does not use all of the information in the probability density of yields, it fully exploits the information in the first and second
Bliss and McCullochKwon (1993) use slightly different filtering procedures; thus the yields they report over periods of overlapping data do not match exactly. I use the yields in McCulloch and Kwon over their entire sample period and use the Bliss (1997) data after February 1991.
conditional moments of the term structure. Thus, QML will capture the ten sion in affine models between fitting conditional means and conditional variances.
Another advantage of QML (which it shares with maximum likelihood and related techniques) is that there is a positive probability that the estimated model could actually generate the observed time series of term structures. This is an important concern in estimating affine term structure models. As the discussion in Section I1 highlighted, there is a tradeoff between fitting the coefficients of variation for expected excess bond returns and fitting the term structure shapes in the data. A model estimated with QML will guar antee that the timet state vector implied by timet yields is in the state vector's admissible space (to avoid a likelihood of zero). By contrast, consider techniques such as Efficient Method of Moments (EMM) that compare Sam ple moments from the data with population moments simulated from the model. These techniques do not require that the estimated term structure model be sufficiently flexible to reproduce the term structure shapes in the data. The parameters of the model in Dai and Singleton (2000), which were estimated with EMM, illustrate this point.
I implement QML following Fisher and Gilles (1996), which contains fur ther details. I assume that at each monthend t, t = l,.. . ,T, yields on n bonds are measured without error. These bonds have fixed times to maturity T~, Yields on k other bonds are assumed to be measured with serially
. . .,T,.
uncorrelated, meanzero measurement error.
Stack the perfectly observed yields in the vector Y, and the imperfectly observed yields in the vector Ft. Denote the parameter vector by O. Given O, Y, can be inverted using (4) to form an implied state vector x,, as in (20).
In (20), H, is an nvector with element i given by A(ri)/ri, and H, is an n x n matrix with row i given by B(T~)/T? The candidate parameter vector is re quired to be consistent with Yt.This is enforced by requiring X, to be in the admissible space for X,, which is equivalent to requiring that the diagonal elements of St in (19c) be real.
Given x,, implied yields for the other Ahbonds can be calculated. Stack them in F,. The measurement error is et = Ft F,. The variancecovariance matrix of the measurement error is assumed to have the following time invariant Cholesky decomposition:
E(et6;) = CC'. (21)
To compute the quasilikelihood value, assume that the oneperiodahead conditional distribution of the state variables is multivariate normal and equal to
The mean and variancecovariance matrix of Xt+, are known (see the Appendix); thus, f,(X,+,IX,) is known. Then the distribution of Yt+, condi tional on Y+ is
Also, assume that the measurement error is jointly normally distributed with distribution f,(e,). The log likelihood of observation t is then
Stationarity is imposed by requiring that the eigenvalues of K are posi tive, allowing fy(Y,IYo) to be set equal to the unconditional distribution of Y,. The estimated parameter vector 0:':is chosen to solve
In the estimation that follows, I assume that the bonds with no measure ment error are those with maturities of 6 months, 2 years, and 10 years. This choice was motivated by the desire to span as much of the term struc ture as possible without assuming that the 3month yield, which exhibits some idiosyncratic behavior, is observed without error. The bonds with mea surement error fill in the gaps in this term structure, with maturities of 3 months, 1year, and 5 years.
D. The Maximization Technique
The QML functions for these models have a large number of local maxima. The most important reason for this is the lack of structure placed on the feedback matrix K. Similar QML values can be produced by very different interactions among the elements of the state vector. Another reason is that the feasible parameter space is not convex for any model with nonconstant volatilities. A feasible parameter vector satisfies the requirement that the diagonal elements of Stare real for all t. Because I use the canonical form of Section I.G, this requirement is satisfied when x,,~2 0 for i 5 m. (Recall that m is the number of state variables that affect the instantaneous vola tility of X,.) Therefore, the requirement imposes m x T restrictions on the parameter vector. The restrictions are nonlinear functions of the parameters and the data. These problems led to the following maximization technique.
Step 1. Randomly generate parameters from a multivariate normal dis tribution with a diagonal variancecovariance matrix. The means and vari ances were set to 'plausible' values.
Step 2. Use (20) to calculate X, for all t.
Step 3. If the parameter vector is not feasible, return to step 1;otherwise
proceed.
Step 4. Use Simplex to determine the parameter vector that maximizes
the QML value.
Step 5. Using the final parameter vector from Step 4 as a starting point, use NPSOL to make any final improvements in the QML value.
This procedure is repeated until Step 5 is reached 1,000 times. For most of the models, there was little improvement in the QML value after the first few hundred iterations.
E. Specification Tests
These specification tests use the fact that QML estimation can be viewed as a GMM estimator. The moments are the first derivatives of the quasi log likelihood function with respect to the parameter vector, resulting in an ex actly identified model. By imposing overidentifying moment conditions we test the adequacy of the model.
E.1. Tests of Nested Models
For m < n, the completely affine model A,(n) is nested in a corresponding essentially affine model. The essentially affine version has an additional n(n m) free parameters corresponding to the bottom n m rows in the matrix A,. We can test the null hypothesis that these free parameters are all equal to zero, using the GMM version of a likelihood ratio test. A textbook discussion is in Greene (1997). Define the column vector ht(0) as the deriv ative of (22) with respect to the parameter vector 0,and define h(0) as the mean of these T vectors. Define the (inverse of) the weighting matrix W, as
Denote the parameter vector for the essentially affine model estimated by QML as 0;. The parameter vector 0, is an alternative vector that imposes the completely affine restriction on A,. Choose it to solve
q = min Th (O,)'W, (0;)h (0,).
01
The results of Hansen (1982) imply that under the null hypothesis, q is distributed as ,y2((n m)n). Similar tests can be used to evaluate other parametric restrictions on the estimated models. These other tests are dis cussed in more detail in Section IV.
E.2. Testing the Covariance Between Forecast Errors and the Term Structure Slope
This test asks whether the yield forecasts produced by the estimated mod els include the information in the slope of the term structure. Given a pa rameter vector O associated with a particular model, the implied state vector x,, is given by inverting yields observed at time t A. The Aperiodahead conditional mean E(X,IX,,) can then be constructed. Given this expected state vector, expected Aperiodahead bond yields and associated forecast errors can also be constructed. We need some notation for forecast errors. Denote by e,,,,, the forecast error realized at time t for a T,maturity bond, where the forecast is made at time t A. The forecast errors for v bonds of different maturities are stacked in the vector e,,,.
If an estimated term structure model does not make systematic forecast errors, forecasts of timet yields made at time t A should have forecast errors uncorrelated with any variable known at time t A. This motivates the specification test. Denote the slope of the yield curve at time t A by st,. If the model is correctly specified,
Equation (25) contains v moments that can be used as overidentifying restrictions in GMM estimation of an affine model. The other moments are standard QML moments, which are the derivatives of (22) with respect to each element of the parameter vector. The weighting matrix is calculated at the QML parameter estimates, which are consistent under the null hypoth esis that the model is correctly specified. Then an analogue to q in (24) is calculated. Again from the results of Hansen (1982), this value is distributed as ,y2(v) under the null hypothesis. The use of overlapping observations in these moment conditions produces sample moments that exhibit serial cor relation. I therefore construct the weighting matrix following Newey and West (1987).
To implement this test, I set A = 112, so that sixmonthahead forecasts are examined. This horizon was chosen arbitrarily. A cursory investigation of other forecast horizons indicated that the results were insensitive to this choice. I used eight lags in the NeweyWest calculation of the weighting matrix; experimentation with similar lag lengths did not materially affect the results. I set v = 3, and formed forecasts for maturities of 6 months, 2 years, and 10 years. (These are the same maturities that are assumed to have no measurement error.) The slope of the term structure is measured by the difference between the 5year bond yield and the 3month bond yield. The first six observations are dropped to account for the length of the fore cast horizon.
Table I1 Statistical Comparison of Estimated Models
Threefactor affine models are estimated with quasi maximum likelihood (QML). The data are monthend yields on zerocoupon bonds with maturities between three months and 10 years, from January 1952 through December 1994. The models differ in the number of factors m that affect the instantaneous variance of yields and in the flexibility of the price of risk parameter ization. Essentially affine models allow the price of risk to vary independently from the in stantaneous variance of yields, while completely affine models do not. "Unrestricted" models impose no restrictions on the parameters other than those required by noarbitrage. "Preferred models drop parameters that contribute little to the models' QML values.
Two specification tests are reported. The first is of the null hypothesis that the model's parameter restrictions are true. For "unrestricted models, the test compares completely affine models to their more general essentially affine counterparts. For "preferred" models, the test compares the preferred model to its unrestricted counterpart. The second tests the null hypoth esis that the sixmonthahead yield forecast errors for bonds of three different maturities are uncorrelated with the slope of the term structure at the time the forecasts are made. Under the null, the test statistics are distributed as X2(n~~mber
of param restrictions) and x2(3), respectively.
Number of  First Test Stat  Second Test Stat  

Model Type  Free Params  QML value  (pvalue)  (pvalue) 
Unrestricted  
Completely  15,171.94  
Completely  15,380.31  
Completely  15,395.74  
Completely  15,396.34  
Essentially  15,196.45  
Essentially  15,392.47  
Essentially  15,396.04  
Preferred  
Completely  
Essentially  
Essentially 
IV. Results
A. Overview
Table I1 reports the QML values for each estimated model. Results for 10 models are shown. The first 7 model specifications are labeled "unrestrict ed." This means that the only parameter restrictions imposed are those im plied by the canonical form. These restrictions are either normalizations or requirements of noarbitrage. To both limit the danger of overfitting and to aid in the interpretation of the parameter estimates, more parsimonious spec ifications are also estimated. They are discussed in Section 1V.B.
There are three main points to take from Table 11. First, models that are better able to produce timevarying volatilities have higher QML values. The A,(3) models (both completely and essentially affine), which have time invariant yield volatilities, have the lowest QML values. As the number of factors that affect volatilities increases from zero through three, QML val ues monotonically rise.
Second, the additional flexibility offered by essentially affine models over completely affine models is important. The first specification test reveals that the completely affine A,(3) and A,(3) models are overwhelmingly re jected by their more general essentially affine counterparts. Recall that as m increases, the flexibility provided by essentially affine models diminishes. We see empirical evidence of this in decreasing values of the first test sta tistic as m increases, culminating in the lack of rejection of the completely affine A2(3) model relative to its essentially affine counterpart.
Third, only the essentially affine model with the greatest flexibility in producing timevarying risk premia can capture the forecasting power of the termstructure slope. According to the second specification test, the yield forecast errors produced by the purely Gaussian essentially affine model (A,(3)) are uncorrelated with the slope of the term structure. For this model, all elements of A, are free. When m = 1,the top row of A, is set to zero. This additional restriction leads to a strong rejection by the second specification test. A similar rejection accompanies the A2(3) essentially affine model. None of the completely affine models pass this second specification test.
We can see from these results the tradeoff between fitting conditional volatilities and producing accurate forecasts of future yields. The essentially affine model with the greatest forecasting power also has the least ability to fit conditional volatilities. An increase in m provides for greater flexibility in fitting conditional variances of yields but also provides for less flexibility (in an essentially affine model) in fitting expected excess returns to bonds. The QML values indicate that the overall goodness of fit of first and second moments is increased by giving up flexibility in forecasting to acquire flex ibility in fitting conditional variances.
B. Parameter Estimates
To limit the size of the paper, I report more detailed information for only three of the models. They are the essentially affine A,(3), A,(3), and com pletely affine A,(3) models. The first is of particular interest because of its forecasting ability, the second illustrates the tradeoff between forecasting ability and fitting conditional variances, while the third is the completely affine model that does the best at forecasting, as measured by the ,y2 statistic on the second specification test.
For each of these models, I estimate a more parsimonious specification by first computing the tstatistics for the unrestricted parameter estimates. I then set to zero all parameters for which the absolute tstatistics did not exceed one and reestimated the models.5 This procedure eliminated five pa rameters from the completely affine A2(3) model and seven parameters from both the essentially affine A0(3) model and the essentially affine A1(3) model. For each preferred model, a joint test of the parameter restrictions is con structed using an analogue to (24). The test statistics and corresponding pvalues are reported in the "First Test Stat" column.6
Parameter estimates for these preferred models are in Tables I11 through
V. To conserve space, parameter estimates for the other models are not re ported in the paper, and are available on request.
Table I11 reports the parameter estimates for the Ao(3) essentially affine model. The canonical form imposes a lower triangular structure on K and imposes a = 1, ,G = 0, KO = 0. Table IV reports the parameter estimates for the A1(3) essentially affine model. One feature of this table deserves high lighting. The parameter (KO), is nonzero, but no standard error is reported. This is the result of two normalizations imposed on the model: 6, = 0, = 0. The normalizations are imposed by setting (KO), and (KO), to the necessary values given K. Other restrictions imposed in the canonical form are a, = k12 = k13 = 0,a2 = a, = Pll = 1, and Pij = 0, i 2 1,j > 1.Finally, Table V reports the parameter estimates for the A2(3) completely affine model. In the canonical form of the A,(3) model, a, = a, = P,, = 0, a, = Pll = P,, = 1, Pij = 0 for i < 3, i # j, and A, = 0. The preferred specification sets P,, = 0 and p,, = 1, so that the second state variable drives the conditional volatil ities of both the second and third state variables. The element (KO), is non zero with no standard error, because 6, = 0 in the canonical model.
C. An Analysis of Forecast Errors
The estimated models, combined with month t bond yields, can be used to construct forecasts of month t + i bond yields. Here we examine the accuracy of these forecasts, both in sample and out of sample. The insample period is January 1952 through December 1994. The outofsample period is January 1995 through December 1998. We focus on bonds with maturities of 6 months, 2 years, and 10 years, and forecast horizons of 3, 6, and 12 months. Forecast accuracy is measured by the root mean squared forecast error (RMSE). In sample RMSEs are reported in Table VI and outofsample RMSEs are re
With the completely affine A,(3) model, the parameter p,, was set to one instead of zero.
The test statistic for the essentially affine A,(3) model suggests a rejection of the preferred model in favor of the unrestricted model. However, the large test statistic appears to be a consequence of approximation errors in numerical computation of the derivative of the log likelihood function with respect to k,,. The Numerical Recipes dfridr routine (a robust method for calculating derivatives and estimates of errors in the derivatives) reported large errors regardless of the initial stepsize. Because the estimate of parameter in the unrestricted model was nearly zero, and setting it to zero had a negligible effect on the QML likelihood function, I set it to zero in the preferred model.
Table I11
Parameter Estimates for the Preferred Essentially
Affine A,(3) Model
The model is defined in equation (19). With this version of the model, a is a vector of ones and both P and KO are identically zero. The matrix C is the Cholesky decomposition V = CC' of the variancecovariance matrix of the crosssectional errors in fitting yields on bonds with matu rities of three months, one year, and five years. Parameters are estimated with QML. Asymptotic standard errors are in parentheses.
Constant Term
Index Number (1)
Parameter 1 2 3
ported in Table VIII. In Tables VII (insample) and IX (outofsample), forecast errors are regressed on the slope of the yield curve to determine whether the models' forecasts capture the forecasting power of the slope.
We need benchmarks to use in evaluating forecast accuracy. The simplest benchmark is a random walk. The month t yield on a Tmaturity bond is used as a forecast of the month t + i yield on a 7maturity bond. The RMSEs associated with this forecast method are reported in the "RW columns of Tables VI and VIII. Note that the tables report different patterns in RMSEs across bonds. In the earlier period, yields were more volatile, with volatility declining with maturity. In the later period, yield volatility was higher at
Table IV
Parameter Estimates for the Preferred Essentially
Affine A,(3) Model
The model is defined in equation (19). With this version of the model, a, = P,, = P,, = 0, a, = a, = pll = 1, and the first row of A, is zero. The matrix C is the Cholesky decomposition V = CC' of the variancecovariance matrix of the crosssectional errors in fitting yields on bonds with maturities of three months, one year, and five years. Parameters are estimated with QML. Asymptotic standard errors are in parentheses.
Constant Term
6, 0.014
(0.005)
Index Number (i)
Parameter 1 2 3
long maturities than at short maturities. Thus, the outofsample period should provide a good test of the robustness of the estimated affine models.
A more sophisticated benchmark uses OLS regressions that predict future changes in yields with the current slope of the term structure. The regres sion is
Table V
Parameter Estimates for the Preferred Completely
Affine A,(3) Model
The model is defined in equation (19). With this version of the model, a, = a, = 0,a, = PI, = p,, = P3, = 1, the remaining elements of the p matrix are zero, and A, is a matrix of zeros. The matrix C is the Cholesky decomposition V = CC' of the variancecovariance matrix of the crosssectional errors in fitting yields on bonds with maturities of three months, one year, and five years. Parameters are estimated with QML. Asymptotic standard errors are in parentheses.
Constant Term
Index Number (i)
Parameter 1 2 3
The parameters of (26) are estimated using insample data. The equation is then used to construct forecasts and forecast errors for both the insample and outofsample periods. The resulting RMSEs are in the columns labeled "OLS" in Table VI and Table VIII. Although the insample RMSE for the regression is guaranteed to be no larger than the random walk RMSE, that is not true out of sample. Indeed, for eight of the nine combinations of ma turity and forecast horizon, the outofsample OLS RMSE exceeds that of the random walk.
The insample parameter estimates from (26) are reported in Table VII in the column labeled "RW." This may seem like a misprint (why aren't they labeled "OLS"?), but recall that Table VII reports the parameter estimates of regressions of forecast errors on the month t slope of the yield curve. With the randomwalk method of forecasting, the regression examined in Table VII is identical to the regression used to produce OLS forecasts. The results
Table VI Comparison of Insample Forecasting Performance
This table reports root mean squared errors (RMSE) for month t forecasts of month t + i bond yields. Eight different forecast methods are compared. The column labeled "RW" (random walk) uses month t yields as forecasts of future yields. The column labeled "OLS" uses a univariate OLS regression to form forecasts, where the dependent variable is change in the yield from t to t + i and the regressor is the month t slope of the yield curve. The final six columns use either completely affine (C.A.) or essentially affine (E.A.) threefactor models to form forecasts. Pre ferred models are restricted versions of unrestricted models. The models differ in the number of factors j that are allowed to affect conditional volatility (AJ(3)).
The regression and affine models are estimated using data from January 1952 through De cember 1994, and the forecasts are produced over the same period (insample forecasts). The slope of the yield curve is the fiveyear zerocoupon yield less the threemonth zerocoupon yield. Bond yields are measured in decimal form (i.e., 0.04 corresponds to four percent/year).
Unrestricted  Preferred  

E. A.  E. A.  
Bond  Forecast  C. A.    C.A.    
Maturity  Horizon  RW  OLS  A,(3)  A,(3)  A,(3)  A,(3)  A,(3)  A,(3)  
6 months  
2 years  
10 years  
6 months  
2 years  
10 years  
6 months  
2 years  
10 years 
document that shortmaturity yields tend to rise and longmaturity yields tend to fall when the slope is steeper than average, although the statistical evidence at the short end is weak. These results correspond to the standard violations of the expectations hypothesis of interest rates.
This violation is also apparent in the behavior of bond yields in the out ofsample period. The "RW column in Table IX reports the results of esti mating (26) from January 1995 through December 1998. The point estimates are typically more negative than their counterparts in Table VII, although the tstatistics are smaller owing to fewer observations.
The final six columns in Table VI through Table IX examine the forecast ing ability of various affine models. The results document that the com pletely affine A,(3) model is a failure at forecasting future interest rates. Table VI reports that in sample, both the unrestricted and preferred speci fications produce forecasts that are worse than those produced by the as sumption that yields follow random walks. This unimpressive performance is mirrored by the performance of the other completely affine models exam ined in this paper. For every estimated model, the assumption that yields follow a random walk produces superior insample forecasts for each of these
Table VII
The Relation Between Insample Forecast Errors and
the Yieldcurve Slope
Various models are used to produce month t forecasts of month t + i bond yields and the corresponding forecast errors are constructed. This table reports parameter estimates from regressions of forecast errors on the month t slope of the yield curve. Six forecast methods are compared. The column labeled "RW (random walk) uses month t yields as forecasts of future yields. Therefore, the forecast error regression is simply a regression of changes in bond yields from t to t + i on the month t yieldcurve slope. The final six columns use either completely affine (C.A.) or essentially affine (E.A.) threefactor models to form forecasts. Preferred models are restricted versions of unrestricted models. The models differ in the number of factors j that are allowed to affect conditional volatility (A,(3)).
The regression and affine models are estimated using data from January 1952 through De cember 1994, and the forecasts are produced over the same period (insample forecasts). The slope of the yield curve is the fiveyear zerocoupon yield less the threemonth zerocoupon yield. Asymptotic tstatistics, in parentheses, are adjusted for generalized heteroskedasticity and moving average residuals.
Unrestricted  Preferred  

E. A.  E. A.  
Bond  Forecast  C. A.  C. A.  
Maturity  Horizon  RW  A,(3)  A0(3)  A,(3)  A,(3)  A0(3)  A,(3) 
6 months  
2 years  
10 years  
6 months  
2 years  
10 years  
6 months  
2 years  
10 years 
maturities and forecast horizons. (These additional results are not reported in any table.)
The regressions reported in Table VII show that the forecast errors of the completely affine A,(3) model are strongly negatively correlated with the slope of the term structure. The parameter estimates are more negative than are the corresponding parameter estimates in the random walk case. The model completely misses the forecasting information in the slope of the term structure. When the term structure is more steeply sloped than usual, the
Table VIII Comparison of Outofsample Forecasting Performance
This table reports root mean squared errors (RMSE) for month t forecasts of month t bond
T z yields. Eight different forecast methods are compared The column labeled "RW (random walk) uses month t yields as forecasts of future yields. The column labeled "OLSuses a univariate OLS regression to form forecasts, where the dependent variable is change in the yield from t to t + i and the regressor is the month t slope of the yield curve. The final six columns use either completely affine (C.A.) or essentially affine (E.A.) threefactor models to form forecasts. Pre ferred models are restricted versions of unrestricted models. The models differ in the number of factors j that are allowed to affect conditional volatility (AJ(3)).
The regression and affine models are estimated using data from January 1952 through De cember 1994, while the forecasts are produced over January 1995 through December 1998 (outofsample forecasts). For each bond there are 48 i forecasts and associated errors. The

slope of the yield curve is the fiveyear zerocoupon yield less the threemonth zerocoupon yield. Bond yields are measured in decimal form (i.e., 0.04 corresponds to four percentlyear).
Unrestricted  Preferred  

E. A.  E. A.  
Bond  Forecast  C. A.      C.A.  
Maturity  Horizon  RbV  OLS  A2(3)  A,(3)  A,(3)  A2(3)  A,(3)  A,(3)  
6 months  
2 years  
10 years  
6 months  
2 years  
10 years  
6 months  
2 years  
10 years 
OLS forecast is that longmaturity yields will fall, but the model forecasts that the yields will rise. Put differently, the model is consistent with the expectations hypothesis, and the observed bond yields are not.
This poor forecasting performance carries over to the outofsample pe riod. Table VIII documents that the unrestricted specification produces fore casts that are inferior to randomwalk forecasts in five of the nine combinations of maturity and forecast horizon. The preferred specification does even worse, producing inferior forecasts for seven of the nine combinations. The point estimates in Table IX confirm that the model's forecasts get the wrong sign of the relationship between the slope of the term structure and future changes in yields.
The essentially affine models produce dramatically better forecasts. The most successful forecasting model, both in sample and out of sample, is the essentially affine, completely Gaussian model. Table VI documents that within the sample, both the unrestricted and preferred A,(3) models outforecast the OLS regressions (and therefore also outforecast the randomwalk assump tion) for each combination of maturity and forecast horizon. Table VIII makes the same point out of sample. Moreover, these forecasts capture the predic
Table IX
The Relation Between Outofsample Forecast Errors and
the Yieldcurve $lope
Various models are used to produce month t forecasts of month t + i bond yields and the corresponding forecast errors are constructed. This table reports parameter estimates from regressions of forecast errors on the month t slope of the yield curve. Six forecast methods are compared. The column labeled "RW (random walk) uses month t yields as forecasts of future yields. Therefore, the forecast error regression is simply a regression of changes in bond yields from t to t + i on the month t yieldcurve slope. The final six columns use either completely affine (C.A.) or essentially affine (E.A.) threefactor models to form forecasts. Preferred models are restricted versions of unrestricted models; model parameters that add little to the model's QML value are set to zero. The models differ in the number of factors j that are allowed to affect conditional volatility (A,(3)).
The regression and affine models are estimated using data from January 1952 through De cember 1994, while the forecasts are produced over January 1995 through December 1998 (outofsample forecasts). For each bond, there are 48 i forecasts and associated errors. The slope of the yield curve is the fiveyear zerocoupon yield less the threemonth zerocoupon yield. Asymptotic tstatistics, in parentheses, are adjusted for generalized heteroskedasticity and moving average residuals.
Unrestricted  Preferred  

E. A.  E. A.  
Bond  Forecast  C. A.  C. A.  
Maturity  Horizon RW A2(3)  A,(3)  A,(3) A2(3)  A,(3)  A,(3)  
6 months  
2 years  
10 years  
6 months  
2 years  
10 years  
6 months  
2 years  
10 years 
tive power of the termstructure slope. In Tables VII and IX, the only evi dence for predictability of forecast errors is in the outofsample forecast errors for 10year bonds at the 12month horizon.
The essentially affine A,(3) model is not quite as successful as the Gauss ian model at forecasting. From Table VI, we see that insample forecasts from both the unrestricted and preferred specifications are typically supe rior to randomwalk forecasts, but outforecast OLS regressions for only half of the maturity/horizon combinations. Moreover, from Table VIII, the fore cast errors are negatively correlated with the slope of the yield curve. The statistical strength of this negative correlation rises as both the bond's ma turity and the forecasting horizon lengthen.
An examination of Table VIII indicates that this essentially affine model performs somewhat better out of sample. Forecasts from the preferred spec ification are superior to randomwalk and OLS forecasts at all maturities and forecast horizons. Nonetheless, Table IX indicates that the model's out ofsample forecast errors are negatively correlated with the slope of the yield curve. Thus, the model misses some of the explanatory power of the term structure slope.
Overall, these results indicate that for the purposes of forecasting, com pletely affine models are largely useless. Even the simplest, most naive rulea random walkdominates the explanatory power of completely affine mod els. A corollary is that we should not use completely affine models to attempt to understand why the expectations hypothesis fails, because the models cannot reproduce this failure.7 By contrast, forecasts from a purely Gaussian essentially affine model dominate naive forecasts.
D. The Predictability of Excess Returns and Volatilities
A few diagrams help shed light on the behavior of these competing models. Figure 1is a graphical summary of the behavior of the preferred essentially affineA,(3) model. Panel A displays instantaneous effects that onestandard deviation shocks to each factor have on the term structure of yields. The three shocks can be interpreted as a level shock (the long dashes), a slope shock (the solid line), and a twist (the short dashes). Panel B displays the (nonexistent) instantaneous effect of these shocks on yield variances.
Panel C displays the effect that these shocks have on bonds' instantaneous expected excess returns (over r,). There are two distinct types of shocks to expected returns. The short dashes correspond to the twist shock in Panel A. This shock has a strong effect on instantaneous expected returns, but it is also very shortlived. (This latter fact cannot be seen in the panel.) Thus, this shock is responsible for highfrequency fluctuations in expected excess returns.
The other type of shock to expected excess returns corresponds to the slope shock in Panel A. It is more persistent (this also cannot be seen in the panel), and thus accounts for more persistent fluctuations in expected returns. The combined effects of these shocks on expected excess returns to 2year bonds are displayed in Panel E. Panel F is the same plot for 10year bonds. These latter panels show that expected excess returns fluctuate sharply and widely around zero. For example, the expected instantaneous excess return in Panel E has a mean of 1.25 percent and a standard deviation of 3.09 percent.
See Dai and Singleton (2001) for a related perspective.
A. Shocks to yields D. Insta~lta~leous
interest rate
Maturity (years) Date
g B. Shocks to variances of yields E. Expected excess return, 2year bond
m

2
X
E
K 2 u a 0
a 5z
g 2 a0
&90 2 4 6 8 1012 1960 1970 1980 1990
Maturity (years) Date
C. Shocks to expected excess returns F. Expected excess return, 10year bond
Matur~ty (years) Date
Figure 1. Summary of the estimated essentially affine A,(3) model. Panels A through C display the instantaneous responses of yields, variances, and expected excess returns (over r,) to onestandarddeviation shocks to each of the three factors. Panels D through F display fitted expected instantaneous returns over the sample period January 1952 through December 1994. Panel D is the instantaneous interest rate. Panels E and F are the instantaneous expected excess returns to the 2year and 10year bonds.
Because this model is so successful at forecasting future yields, it is worth a more careful examination. An intuitive way to interpret shocks to bond yields is to decompose the shocks into shocks to expected future shortterm interest rates and shocks to expected excess returns. This decomposition is straightforward; thus I will not discuss it in detail here. Instead, I will sim ply summarize the results.
A positive level shock corresponds to an immediate, nearpermanent in crease in shortterm interest rates. The halflife of the shock to shortterm interest rates is more than 11years. Because the shock does not substan tially alter investors' required excess returns to bonds, shortmaturity and longmaturity bond yields respond in the same way to this shock.
A positive slope shock corresponds to an immediate increase in shortterm interest rates that lasts about as long as a business cycle. The halflife of the shock is four years. Because shortterm interest rates are expected to de cline over time, the shock lowers the slope of the term structure. The shock also lowers expected excess returns to bonds by affecting the price of risk vector. We can see this in the parameters of A, in Table 111.An increase in the first factor (the slope factor) affects the price of risk of the third factor (the level factor) through element (3,l) of A,. This decrease in expected re turns further decreases the slope of the term strucure because longer maturity bond returns are more sensitive than shortermaturity bond returns to level shocks, and thus to the price of risk of level shocks.
Twists are very similar to the 'Ttnfactor in the twofactor example dis cussed in Section 11. A twist shock has basically no effect on current or fu ture shortterm interest rates. Instead, the shock changes investors' required excess returns to bonds by affecting the price of risk associated with the level and slope factors. The halflife of such a shock is less than three months. We can call this a "flight to quality" shock. Investors experience shortlived periods of unwillingness to hold risky Treasury instruments, thus driving expected excess bond returns higher.
Figure 2 contains information about the preferred essentially affine A,(3) model. Panel A displays a level shock, a slope shock, and a twist shock. The solid line is the level shock, and it affects the conditional variance of yields, as shown in Panel B. The longdashed line is the twist shock, and in Panel C, we see its strong effect on expected excess returns. However, Panel C also indi cates that the other two shocks have little effect on expected excess returns. The net effect is that in Panels E and F, the fluctuations in expected excess returns are less volatile than the fluctuations in the corresponding panels in Figure 1. For example, the expected instantaneous excess return in Panel E has a mean of 1.90 percent and a standard deviation of 1.85 percent.
Why does a shock to the slope affect expected excess returns in Figure 1 but not in Figure 2? The answer is that the channel that operates in the model underlying Figure 1is unavailable in the model underlying Figure 2. Panel C in Figure 1reflects a relationship between shocks to the slope and shocks to the price of risk of level shocks. These crossfactor relationships are more limited in the essentially affine A,(3) model. In the canonical form, the first factor drives conditional volatilities; thus its price of risk cannot be affected by any other factors. Figure 2 indicates that this first factor is the level factor; shocks to the slope cannot affect its price of risk. Therefore, this model produces poorer forecasts of future bond yields than does the essen tially affine A,(3) model.
Figure 3 displays the same panels for the preferred completely affine A2(3) model. The model generates a richer pattern of time variation in volatilities than do the other two models. The cost of these more accurate measures of volatility is an inability to fit expected excess returns. Expected excess re turns in Panels E and F are always positive, never large, and not volatile. For example, the expected instantaneous excess return in Panel E has a mean of 0.79 percent and a standard deviation of 0.41 percent. Moreover, these expected excess returns roughly track the instantaneous interest rate displayed in Panel D. Because higher shortterm rates typically correspond to lower slopes, the figure indicates that expected excess returns move in versely with the slope of the yield curve, but this is counterfactual.
A. Shocks to yields D. Instantaneous interest rate
w
..
6,
0,E  a  in  

0  
0  2  4  681012  1960  1970  1980  1990  
Matunty (years)  Date  
B. Shocks to variances of yields  E. Expected excess return, 2year bond  
M 
6, a 6, o
w
E
G
1960 1970 1980 1990 Maturity (years) Date
C. Shocks to expected excess returns F. Expected excess return, 10year bond
Maturity (years) Date
Figure 2. Summary of the estimated essentially affine A,(3) model. Panels A through C display the instantaneous responses of yields, variances, and expected excess returns (over I.,) to onestandarddeviation shocks to each of the three factors. Panels D through F display fitted expected instantaneous returns over the sample period January 1952 through December 1994. Panel D is the instantaneous interest rate. Panels E and F are the instantaneous expected excess returns to the 2year and 10year bonds.
The results discussed in this section indicate that the completely affine A2(3) model fails to reproduce the behavior of expected excess returns to Treasury bonds. The same conclusion holds for the other completely affine models estimated in this paper that are not discussed in detail here. The models systematically fail to capture the large fluctuations in expected ex cess returns to bonds. Essentially affine models do a better job of reproduc ing the behavior of expected excess returns, although the magnitude of the improvement is inversely related to the ability of the models to fit the time variation in conditional variances of yields.
V. Concluding Comments
Recent term structure research has concentrated on what I call completely affine models. This paper documents that completely affine models do not forecast future yields well over the nearly 50year period examined here.
A. shocks to yields D. Instantaneous interest rate
;'
" 0
a r;r
0
0246 81012
Maturity (years) Date
3 B. Shocks to variances of yields E. Expected excess return, 2year bond
2
izri icp
x b
& 2
+ ._E E?
E g
& 024681012 1960 1970 1980 1990
Maturity (years) Date
C Shocks to expected excess returns F. Expected excess return, 10year bond
!2Fi!:pi
B 0 
5 O w
E
5 PJ
d
a o a

0 2 4 6 8 1012 1960 1970 1980 1990
Maturity (years) Date
Figure 3. Summary of the estimated completely affine A,(3) model. Panels A through C display the instantaneous responses of yields, variances, and expected excess returns (over r,) to onestandarddeviation shocks to each of the three factors. Panels D through F display fitted expected instantaneous returns over the sample period January 1952 through December 1994. Panel D is the instantaneous interest rate. Panels E and F are the instantaneous expected excess returns to the 2year and 10year bonds.
They consistently underestimate future returns to bonds when the term struc ture is more steeply sloped than usual; put differently, these models do not reproduce the wellknown failure of the expectations hypothesis.
Essentially affine models generalize completely affine models. They allow greater flexibility in fitting variations in the price of interest rate risk over time, while retaining the affine timeseries and crosssectional properties of bond prices. One of the essentially affine models investigated in this paper the pure Gaussian modelgenerates reasonable forecasts of future yields, in the sense that the predictive power of the term structure is subsumed within the model's forecasts.
The forecast accuracy of this Gaussian model allows us to properly inter pret the usual level, slope, and twist yieldcurve factors in terms of their predictions for future shortterm interest rates and excess returns to longer term bonds. Level shocks correspond to nearpermanent changes in interest rates and only minimal changes in expected excess returns. Slope shocks correspond to businesscyclelength fluctuations in both interest rates and expected excess returns to bonds, while twist shocks correspond to short lived "flight to quality" variations in expected excess returns. In other words, twist shocks do not affect current or expected future shortterm interest rates; they are pure shocks to risk premia.
Essentially affine models are not magic bullets. The models cannot cap ture time variation in conditional variances without giving up part of their flexibility in fitting time variation in the price of interest rate risk. It re mains to be seen whether an essentially affine model can be constructed that reproduces the time variation observed in both the conditional vari ances of yields and expected returns to bonds.
Appendix
This appendix gives closedform representations for first and second con ditional moments of a state vector that follow the affine process of (12a) and (2). The results are an application (and a specialization) of the results in Fisher and Gilles (1996).
Assume that K can be diagonalized, or
K = NDNP1, D diagonal. (All
The diagonal elements of D are denoted di,. . .,d,. A discussion of computing moments when K cannot be diagonalized is in Fisher and Gilles (1996).
The approach taken here is to compute the first and second conditional moments of a linear transformation of X,. The transformation is chosen so that the feedback matrix K is diagonal under the transformation. The linear transformation is then reversed to calculate the conditional moments of X,. Define
Then the dynamics of X: are, from (12a), (2), (Al), and (A2),
where
We now calculate the first and second moments of XI:. Some notation is helpful. If Z is an nvector, the n x n diagonal matrix in which element (i, i) equals Zi is denoted diag(Z). If Z is a diagonal matrix, the diagonal matrix in which element (i,i) equals eZ~l is denoted eZ. Finally, the nvector P.i is column i of p.
A. Conditional Mean
The expectation of X; conditional on X; is given by
Because e D(Tt) is diagonal, this expectation can also be simply expressed elementbyelement:
Another useful way to express (A4) is by separating the terms that depend on Xf from the terms that do not:
Given this conditional mean of X;, we reverse the transformation to ex press the conditional mean of XT:
Note that the conditional mean of XT could be expressed directly in terms of the parameters of (12a); no transformation into X,'; is required, because the above expression is equivalent to
where eK(Tt) is the fundamental matrix associated with K(T t). The value of the approach taken here is that (A4') is used in determining the conditional variancecovariance matrix of Xt .
B. Conditional Variance
The matrix Z"S:S:'C";' is the instantaneous variancecovariance matrix of the transformed state vector. We can write this as
where Go = C"diag(u"')C'" and the n x n matrices Giare defined as
[C:"iag(P:") C:::'] D 1 . efine the n x n matrix F(t,s) as
This matrix is the instantaneous variancecovariance matrix of X:, but evaluated at the expectation of Xz (conditional on timet information) instead of at the true value of X:.Using (A4'),this matrix can be expressed as
Fisher and Gilles (1996) show the conditional variance of X$ can be written as
Substituting (A6) into (A7) produces (A8):
Iff (j,k) maps (j,k) into the scalar value f, the notation { f (j,k)) denotes the matrix with element (j, h)given by f (j, k). The conditional variance can then be written as
Integrating (A9) produces (A10):
Note that by collecting terms, the variancecovariance matrix in (A10) can be rewritten in terms of the individual elements of X; as in (All):
The n X n matrices bi, i = 0,.. . ,n depend on the horizon T t. We now calculate the conditional variance of X, using the notation of (All). Since
we have
REFERENCES
Andersen, Torben G., and Jesper Lund, 1997, Estimating continuoustime stochastic volatility models of the shortterm interest rate, Journal of Econometrics 77, 343377. Bliss, Robert R., 1997, Testing term structure estimation methods, Advances in Futzrres and Options Research 9, 197231. Chacko, George, 1997, Multifactor interest rate dynamics and their implications for bond pric ing, Working paper, Harvard University.
Chan, K. C., G. Andrew Karolyi, Francis A. Longstaff, and Anthony B. Sanders, 1992, An em pirical comparison of alternative models of the shortterm interest rate, Journal of Finance 47, 12091227.
Cox, John C., Jonathan E. Ingersoll, and Stephen A. Ross, 1985, A theory of the term structure of interest rates, Econometrica 53, 385407. Dai, Qiang, and Kenneth J. Singleton, 2000, Specification analysis of affine term structure models, Journal of Finance 55, 19431978.
Dai, Qiang, and Kenneth J. Singleton, 2001, Expectation puzzles, timevarying risk preinia, and dynamic models of the term structure, Working paper, Stanford University; forthcom ing, Journal of Financial Economics.
Duarte, Jefferson, 2000, The relevance of the price of risk in affine term structure models, Working paper, University of Chicago. Duffie, Darrell, and Rui Kan, 1996, A yieldfactor model of interest rates, Mathematical Fi nance 6, 379406.
Duffie, Darrell, Jun Pan, and Kenneth Singleton, 2000, Transform analysis and asset pricing for affine jumpdiffusions, Ecolzometrica 68, 13431376. Fama, Eugene F., and Robert R. Bliss, 1987, The information in longmaturity forward rates, American Economic Review 77, 680692. Fama, Eugene F., and Kenneth R. French, 1989, Business conditions and expected returns on stocks and bonds, Journal of Financial Economics 25, 2349. Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors in the returns on stocks and bonds, Journal of Financial Economics 33, 356. Fisher, Mark, 1998, A simple model of the failure of the expectations hypothesis, Working paper, Federal Reserve Board. Fisher, Mark, and Christian Gilles, 1996, Estimating exponentialaffine models of the term structure, Working paper, Federal Reserve Board. Greene, William H., 1997, Econometric Analysis, 3rd ed. (PrenticeHall, Upper Saddle River, NJ). Hansen, Lars Peter, 1982, Large sample properties of generalized method of moment estima tors, Econometrica 50, 10291053. Litterman, Robert, and Jose Scheinkman, 1991, Common factors affecting bond returns, Jozrrnu1 of Fizecl Income 1, 5461. McCulloch, J. Huston, and HeonChul Kwon, 1993, U.S. terin structure data, 19471991, Work ing paper 936, Ohio State University.
Newey, Whitney K., and Kenneth D. West, 1987, A simple, positive semidefinite, heteroske dasticity and autocorrelation consistent covariance matrix, Econometrica 55, 703708. Singleton, Kenneth J., 2001, Estimation of affine asset pricing models using the empirical
characteristic function, Journal of Econometrics 102, 111141.
Vasicek, Oldrich, 1977, An equilibrium characterization of the terin structure, Journal of Fi nancial Economics 5, 177188.
Comments