Financial Institutions Center

Forecasting the Term Structure of Government Bond Yields
by Francis X. Diebold Canlin Li 02-34

The Wharton Financial Institutions Center

The Wharton Financial Institutions Center provides a multi-disciplinary research approach to the problems and opportunities facing the financial services industry in its search for competitive excellence. The Center's research focuses on the issues related to managing risk at the firm level as well as ways to improve productivity and performance. The Center fosters the development of a community of faculty, visiting scholars and Ph.D. candidates whose research interests complement and support the mission of the Center. The Center works closely with industry executives and practitioners to ensure that its research is informed by the operating realities and competitive demands facing industry participants as they pursue competitive excellence. Copies of the working papers summarized here are available from the Center. If you would like to learn more about the Center or become a member of our research community, please let us know of your interest.

Franklin Allen Co-Director

Richard J. Herring Co-Director

The Working Paper Series is made possible by a generous grant from the Alfred P. Sloan Foundation

Forecasting the Term Structure of Government Bond Yields
Francis X. Diebold
University of Pennsylvania and NBER

Canlin Li
University of California, Riverside

First Draft, December 2000 This Draft/Print: August 6, 2002

Copyright © 2000-2002 F.X. Diebold and C. Li. This paper is available on the World Wide Web at http://www.ssc.upenn.edu/~diebold and may be freely reproduced for educational and research purposes, so long as it is not altered, this copyright notice is reproduced with it, and it is not sold for profit.

Abstract: Despite powerful advances in yield curve modeling in the last twenty years, comparatively little attention has been paid to the key practical problem of forecasting the yield curve. In this paper we do so. We use neither the no-arbitrage approach, which focuses on accurately fitting the cross section of interest rates at any given time but neglects time-series dynamics, nor the equilibrium approach, which focuses on time-series dynamics (primarily those of the instantaneous rate) but pays comparatively little attention to fitting the entire cross section at any given time and has been shown to forecast poorly. Instead, we use variations on the Nelson-Siegel exponential components framework to model the entire yield curve, period-by-period, as a three-dimensional parameter evolving dynamically. We show that the three time-varying parameters may be interpreted as factors corresponding to level, slope and curvature, and that they may be estimated with high efficiency. We propose and estimate autoregressive models for the factors, and we show that our models are consistent with a variety of stylized facts regarding the yield curve. We use our models to produce term-structure forecasts at both short and long horizons, with encouraging results. In particular, our forecasts appear much more accurate at long horizons than various standard benchmark forecasts. Finally, we discuss a number of extensions, including generalized duration measures, applications to active bond portfolio management, and arbitrage-free specifications.

Acknowledgments: The National Science Foundation and the Wharton Financial Institutions Center provided research support. For helpful comments we are grateful to Dave Backus, Rob Bliss, Michael Brandt, Todd Clark, Qiang Dai, Ron Gallant, Mike Gibbons, David Marshall, Monika Piazzesi, Eric Renault, Glenn Rudebusch, Til Schuermann, and Stan Zin, as well as seminar participants at Geneva, Georgetown, Wharton, the European Central Bank, and the National Bureau of Economic Research. We, however, bear full responsibility for all remaining flaws.

1. Introduction The last twenty-five years have produced major advances in theoretical models of the term structure as well as their econometric estimation. Two popular approaches to term structure modeling are no-arbitrage models and equilibrium models. The no-arbitrage tradition focuses on perfectly fitting the term structure at a point in time to ensure that no arbitrage possibilities exist, which is important for pricing derivatives. The equilibrium tradition focuses on modeling the dynamics of the instantaneous rate, typically using affine models, after which yields at other maturities can be derived under various assumptions about the risk premium.1 Prominent contributions in the no-arbitrage vein include Hull and White (1990) and Heath, Jarrow and Morton (1992), and prominent contributions in the affine equilibrium tradition include Vasicek (1977), Cox, Ingersoll and Ross (1985), and Duffie and Kan (1996). Interest rate point forecasting is crucial for bond portfolio management, and interest rate density forecasting is important for both derivatives pricing and risk management.2 Hence one wonders what the modern models have to say about interest rate forecasting. It turns out that, despite the impressive theoretical advances in the financial economics of the yield curve, surprisingly little attention has been paid to the key practical problem of yield curve forecasting. The arbitrage-free term structure literature has little to say about dynamics or forecasting, as it is concerned primarily with fitting the term structure at a point in time. The affine equilibrium term structure literature is concerned with dynamics driven by the short rate, and so is potentially linked to forecasting, but most papers in that tradition, such as de Jong (2000) and Dai and Singleton (2000), focus only on in-sample fit as opposed to out-of-sample forecasting. Moreover, those that do focus on out-of-sample forecasting, notably Duffee (2002), show that the models forecast miserably. In this paper we take an explicit out-of-sample forecasting perspective, and we use a flexible modeling approach in hope of superior forecasting performance. We use neither the no-arbitrage approach nor the equilibrium approach. Instead, we use the Nelson-Siegel (1987) exponential components framework to distill the entire yield curve, period-by-period, into a three-dimensional parameter that evolves dynamically. We show that the three time-varying parameters may be interpreted as factors. Unlike factor analysis, however, in which one estimates both the unobserved factors and the
1

The empirical literature that models yields as a cointegrated system, typically with one underlying stochastic trend (the short rate) and stationary spreads relative to the short rate, is similar in spirit. See Diebold and Sharpe (1990), Hall, Anderson, and Granger (1992), Shea (1992), Swanson and White (1995), and Pagan, Hall and Martin (1996). For comparative discussion of point and density forecasting, see Diebold, Gunther and Tay (1998) and Diebold, Hahn and Tay (1999). 1
2

showing that the three coefficients in the Nelson-Siegel curve may be interpreted as latent level. and arbitrage-free specifications. and that our approach. Chen (1996). Squassi and Weir (1995). and we also show how it can replicate a variety of stylized facts about the yield curve. We proceed as follows. We propose and estimate autoregressive models for the factors. Particularly relevant are the three-factor models of Balduzzi.. we impose a particular functional form on the factor loadings.factor loadings. although related. which interprets and extends earlier work in ways linked to recent developments in multifactor term structure modeling. slope and curvature. and we then forecast the yield curve by forecasting the factors. Finally. and examining out-of-sample forecasting performance. is indeed very different. Foresi and Sundaram (1996). and we introduce a novel twist of interpretation. In section 2 we provide a detailed description of our modeling framework. slope and curvature. i. In section 4 we conclude and discuss a number of variations and extensions that represent promising directions for future research. but also lets us interpret the estimated factors as level. We also argue that the nature of the factors and factor loadings implicit in the Nelson-Siegel model make it potentially consistent with various stylized empirical facts about the yield curve that have been cataloged over the years. de Jong (2000). and let yt(J) denote its continuously-compounded zero-coupon nominal yield 2 . generalized duration measures. Our results are encouraging. slope and curvature factors. We argue that the well-known Nelson-Siegel (1987) curve is well-suited to our ultimate forecasting purposes. slope and curvature. Bliss (1997a. in particular. Dai and Singleton (2000). whose three factors are interpreted in terms of level. Das. Modeling and Forecasting the Term Structure I: Methods Here we introduce the framework that we use for fitting and forecasting the yield curve. Brandt and Yaron (2001) and Duffee (2002). Doing so not only facilitates highly precise estimation of the factors. including state-space modeling and optimal filtering. de Jong and Santa-Clara (1999). and especially the Andersen-Lund (1997) model with stochastic mean and volatility. estimating the models. motivated by our interpretation of the Nelson-Siegel model as a three-factor model of level. our models produce one-year-ahead forecasts that are strikingly more accurate than standard benchmarks. In section 3 we proceed to an empirical analysis. 1997b). for now. We will subsequently discuss related work in greater detail. Fitting the Yield Curve Let Pt(J) denote the price of a J-period discount bond. we contrast it to various multi-factor models that have appeared in the literature. applications to active bond portfolio management. 2.e. describing the data. the present value at time t of $1 receivable J periods ahead. suffice it to say that little of it considers forecasting directly. Closely related work includes the factor models of Litzenberger.

yield curves. Hence the VasicekFong model is more successful at fitting yield curves with flat long ends. because its estimation requires iterative nonlinear optimization. but its lack of restrictions may be a vice rather than a virtue for forecasting. diverges at long maturities instead of converging to zero.” begins with the unsmoothed Fama-Bliss piecewise linear curve. which we discuss in detail below. Pt(J) ' e ) &J y t(J) . In practice. who model the discount curve with a cubic spline. Smoothed Fama-Bliss effectively does so. A third approach to yield curve fitting is due to Fama and Bliss (1987). which can be conveniently estimated by least squares. who develop an iterative method for piecewise-linear fitting of forward rate curves. A second approach is due to Vasicek and Fong (1982). and even if it could be done it might lead to poor forecasts due to overfitting. Given the yield curve or forward rate curve. who fit exponential splines to the discount curve. One popular approach to yield curve fitting is due to McCulloch (1975) and McCulloch and Kwon (1993). Instead. From the yield curve we obtain the discount curve. and then smooths using the Nelson-Siegel (1987) model.to maturity. Hence such curves provide a poor fit to yield curves that are flat or have a flat long end. Unsmoothed Fama-Bliss appears accurate and unrestrictive. using a negative transformation of maturity instead of maturity itself. discount curves and forward rate curves are not observed. and from the discount curve we obtain the instantaneous (nominal) forward rate curve. Jmt 0 J or ft(J) ' yt(J) % Jyt (J) . we want to distill the entire term structure into just a few parameters. but one 3 ) . ft(J) ' &Pt (J) / Pt(J) . which implies that the zero-coupon yield is an equally-weighed average of forward rates. The relationship between the yield to maturity and the forward rate is therefore 1 yt(J) ' f (u)du . and it can be hard to restrict the forward rates to be positive. which ensures that the forward rates and zero-coupon yields converge to a fixed limit as maturity increases. however. “smoothed Fama-Bliss. The fitted discount curve. which requires an exponentially decreasing discount function. however. because it’s not clear how to extrapolate a nonparametrically-fit curve. It has problems of its own.” A natural extension. we can price any coupon bond as the sum of the present values of future coupon and principal payments. sometimes called “unsmoothed Fama-Bliss. Instead they are estimated from observed prices of bonds by interpolating for missing maturities and/or smoothing to reduce the impact of noise.

We now proceed to examine the Nelson-Siegel approach in greater detail. and then decays to zero (and thus is not long-term). 4 . a constant that does not decay to zero in the limit. $2 and $3 in the Nelson-Siegel model as three latent factors. hence it may be viewed as &8 J 8tJ &8 J 1&e t a short-term factor. a function that starts at 1 but decays monotonically and quickly to 0. Nelson-Siegel is a three-component exponential approximation to the yield curve. which starts at 0 (and is thus not short-term). yt(J) ' $1t % $2t 1&e 8tJ &8tJ &8tJ % $3t 8tJe &8tJ . The factor loading in the Vasicek (1977) model has exactly the same form. and it is from the class of functions that are solutions to differential or difference equations.may as well then go ahead and fit Nelson-Siegel to the raw term structure data. as extended by Siegel and Nelson (1988). for example. as we shall demonstrate. while large values of 8t produce fast decay and can better fit the curve at short maturities. The loading on $1t is 1. Courant and Hilbert (1953). easy to estimate by least squares. % $3t &8 J 1&e t &e t . 8tJ &8 J The Nelson-Siegel forward rate curve can be viewed as a constant plus a Laguerre function. We work with the Nelson-Siegel model because of its ease of interpretation and its parsimony. We interpret $1 .4 The loading on $3t is &e t . small values of 8t produce slow decay and can better fit the curve at long maturities. which is a polynomial times an exponential decay term and is a popular mathematical approximating function. rather than to the FamaBliss term structure – which is precisely what we do. 8tJ increases. The Nelson-Siegel Yield Curve and its Interpretation Nelson and Siegel (1987). Bliss (1997b) compares the different yield curve fitting methods and finds that the Nelson-Siegel approach performs admirably. This brings us to a fourth approach to yield curve fitting. which proves very useful for our purposes.3 The parameter 8t governs the exponential decay rate. work with the instantaneous forward rate curve. The loading on $2t is &8 J 1&e t . has a discount function that begins at one at zero horizon and approaches zero at infinite horizon. as appropriate. which promote simplicity of modeling and accuracy of forecasting. hence it may be viewed as a long-term factor. ft(J) ' $1t % $2t e which implies the yield curve. where 8t is a mean reversion coefficient. due to Nelson and Siegel (1987). It is parsimonious. hence it may be viewed as a medium-term 3 4 See.

factor.78$2t%. as the loading is identical at all maturities.5 The loading on $3t is maximized at a maturity of approximately three years. We plot the three factor loadings in Figure 1.00053$2t%. Foresi and Sundaram (1996) and Chen (1996) impose the restrictions that the instantaneous yield is an affine function of only two of the three state variables. that the instantaneous yield depends on both the level and slope factors. slope and curvature. short-term and medium-term. In particular.06$3t when 8t = 0. In particular. In particular. The short-term factor $2t is closely related to the yield curve slope. Finally. which we define as the tenyear yield minus the three-month yield. In particular. which load minimally on it.6 An important insight is that the three factors. who estimated loadings via a statistical factor analysis. It is interesting to note. In our subsequent empirical work. the unit loading on the first factor is imposed from the beginning. 2yt(24)&yt(3)&yt(120) ' . Now that we have interpreted Nelson-Siegel as a three-factor of level.0609 without significantly degrading the goodness of fit of the time series of fitted term structures. Several other models have the same implication. Factors are typically not uniquely identified in factor analysis. for example.0609. governs the yield curve level. which potentially enables us to estimate the other factors more efficiently.37 $3t when 8t =0. Alternatively. a property shared by the Andersen-Lund (1997) three-factor non-affine model. Dai and Singleton (2000) show that the three-factor models of Balduzzi. yt(120)&yt(3) ' &. one can easily verify that yt(4)'$1t . moreover. which is exactly equal to &$2t . Alternatively.0609. We do so throughout this paper. define the yield curve slope as yt(4)&yt(0) . note that an increase in $1t increases all yields equally. Bliss (1997a) rotates the first factor so that its loading is a vector of ones. and we will provide details in due course. with 8t =0. Das. which we have thus far called long-term. thereby changing the level of the yield curve. the medium-term factor $3t is closely related to the yield curve curvature. moreover. thereby increasing yield curve curvature. slope and curvature. which is highly related yet distinct. note that an increase in $3t will have little effect on very short or very long yields. it is appropriate to contrast it to Litzenberger. which we define as twice the two-year yield minus the sum of the ten-year and three-month yields. Some authors such as Frankel and Lown (1994). may also be interpreted in terms of the aspects of the yield curve that they govern: level. because yt(0) ' $1t % $2t . which load more heavily on it. In our approach. because the short rates load on $2t more heavily. thereby changing the slope of the yield curve. we find that we can fix 8t = 0. 5 6 5 .0609. We have seen that $1t governs the level of the yield curve and $2t governs its slope. Alternatively. note that an increase in $2t increases short yields more than long yields. Squassi and Weir (1995). The long-term factor $1t . The factor loading plots also look very much like those obtained by Bliss (1997a). but will increase medium-term yields.

a sum of factors in the yield curve will not be a sum in the discount curve. Ours is &8 J &8 J 1&e t preferable. so if we were to interpret b2 and b3 as factors. the average yield curve is the yield curve corresponding to the average values of $1t . operationally. Finally. however. and e t 8tJ have similar monotonically decreasing shape.. Obviously the Nelson-Siegel factorization matches ours with b1t ' $1t . who used yt(J) ' b1t % b2t 1&e 8tJ &8tJ & b3t e &8tJ . including upward sloping. 6 . it is hard to provide intuitive interpretations of the factors in the original Nelson-Siegel framework. in closing this sub-section. in principle. Stylized Facts of the Yield Curve and the Three-Factor Model’s Potential Ability to Replicate Them A good model of yield curve dynamics should be able to reproduce the historical stylized facts concerning the average shape of the yield curve. both we and Litzenberger et al. for example. for reasons that we are now in a position to appreciate. $2t and $3t . In the multi-factor case. then their loadings would be forced to be very similar.First. it is difficult to estimate the factors precisely. although Litzenberger et al. (2) The yield curve assumes a variety of shapes through time. the strong persistence of yields and weak persistence of spreads. conceptually. Let us consider some of the most important stylized facts and the ability of our model to replicate them. however. First. b2t ' $2t%$3t . perhaps because term premia may not be adequately captured by affine models. model the discount curve Pt(J) using exponential components and we model the yield curve yt(J) using exponential components. because the high coherence in the factors produces multicolinearity. It is certainly possible in principle that it may be increasing and concave. the yield curve is a log transformation of the discount curve because yt(J) ' &log Pt(J)/J . so the two approaches are equivalent in the one-factor case. (1) The average yield curve is increasing and concave. Duffee (2002). which creates at least two problems. First. however. it is worth noting that what we have called the “Nelson-Siegel curve” is actually a different factorization than the one originally advocated by Nelson and Siegel (1987). Second. and so on. do not interpret parameters directly as factors. and second. and b3t ' $3t . It is not easy for a parsimonious model to accord with all such facts. the variety of shapes assumed at different times. provide novel interpretations of the parameters of fitted curves. shows that multi-factor affine models are inconsistent with many of the facts. Litzenberger et al. so there is generally no simple mapping between the approaches. In our framework.

In our framework. whereas the long end depends only on $1t . The Data We use end-of-month price quotes (bid-ask average) for U.S. and less persistent spread dynamics would correspond to weaker persistence of $2t . Overall. We then calculate the yields by averaging the forward rates.” exactly price the included bonds. We begin by introducing the data. after which we model and forecast the extracted level. If $1t is the most persistent factor. this is reflected in factor loadings: the short end depends positively on both $1t and $2t . Whether and how often it does depends upon the variation in $1t . slope and curvature components. we compute the forward rate necessary to price successively longer maturity bonds. and bonds with special liquidity problems (notes and bonds with less than one year to maturity. The yield curve in our framework can assume all of those shapes. given the yields fitted to previously included issues. Although most of our analysis does not require the use of fixed maturities. The resulting yields.downward sloping. then long rates will be more persistent than short rates. Modeling and Forecasting the Term Structure II: Empirics In this section. (3) Yield dynamics are persistent. In our framework. long rates depend only on $1t . humped. and inverted humped. taken from the CRSP government bonds files. eliminating bonds with option features (callable and flower bonds). doing so greatly 7 We thank Rob Bliss for providing us with the computer programs and data. from January 1985 through December 2000. with many of the key stylized facts of yield curve behavior. We then use the Fama-Bliss (1987) bootstrapping method to compute raw yields recursively from the filtered data. we filter the data before further analysis. it seems clear that our framework is consistent. Whether principle accords with practice is an empirical matter. (4) The short end of the yield curve is more volatile than the long end.7 At each step. 3. we estimate and assess the fit of the three-factor model in a time series of cross sections. and spread dynamics are much less persistent. $2t and $3t . 7 . Persistent yield dynamics would correspond to strong persistence of $1t . Following Fama and Bliss (1987). which we call “unsmoothed Fama-Bliss. (5) Long rates are more persistent than short rates. at least in principle. Treasuries. and bills with less than one month to maturity). to which we now turn.

spread and curvature. That is why affine models don’t fit the data well. or 34 days to maturity. The earliermentioned upward sloping pattern. and graphs of their sample autocorrelations to a displacement of sixty months. Hence we pool the data into fixed maturities. Because not every month has the same maturities available. corr(level. 24. but nevertheless apparent. 84. 36. One can also see that the distributions of yields around their medians tend to be asymmetric. curvature)=-0. as well as the yield curve level. 60. 48. and 120 months. In Table 1.simplifies our subsequent forecasting exercises. and corr(slope.13. that the level (120month yield) is highly persistent but varies only moderately relative to its mean. The very high persistence of the level. 9. we don’t use it in the analysis. curvature)=0. 32. 6 months. See also Duffee (1996) for additional discussion. 31. they can’t generate such high variability and quick mean reversion in curvature. Similarly we obtain data for maturities of 3 months. In Figure 2 we provide a three-dimensional plot of our term structure data. corr(level. 12. 72. The large amount of temporal variation in the level is visually apparent. 108. In Figures 3 and 4 we highlight and expand upon certain of the facts revealed in Table 1. Due to potential problems of idiosyncratic behavior with the 1-month bill. is apparent. that the long rates less volatile and more persistent than short rates. slope)=0.4375 days. In Figure 3 we display time-series plots of yield level. 96. 30. The variation in slope and curvature is less strong. that the slope is less persistent than any individual yield but quite highly variable relative to its mean.40. 21.4375 days to maturity. slope and curvature defined above.9 The various yields. as is the presence of a stochastic cycle in the slope as evidenced by its oscillating autocorrelation function. 33. etc. where a month is defined as 30. Hence we focus on them now in some detail. that level. 6. In particular. we present descriptive statistics for the monthly yields. with a long right tail. and the curvature is the least persistent of all factors and the most highly variable relative to its mean. we linearly interpolate nearby maturities to pool into fixed maturities of 3. slope and curvature are not highly correlated with each other. because it will be relevant for our future modeling choices.8 Although there is no bond with exactly 30. with long rates less volatile than short rates.38. will play a prominent role in the sequel. 15. each month there are many bonds with either 30. 8 10 9 8 . We checked the derived dataset and verified that the difference between it and the original dataset is only one or two basis points. It is clear that the average yield curve is upward sloping. In Figure 4 we display the median yield curve together with pointwise interquartile ranges. 18. moderate persistence of the slope and comparatively weak persistence of the curvature are apparent.10 It is worth noting.

we decided to fix 8t =0. for each month t. with and without 8t fixed. that is. but first let us discuss the fitted values of 8t .0609. In the top panel. and that recent theoretical developments suggest that 8t should in fact be constant (as we will discuss subsequently). yt(J) ' $1t % $2t 1&e 8tJ &8tJ % $3t &8 J 1&e t &e t . $2t. we plot the three-factor RMSE over time (averaged over maturity). Although there is variation over time in the estimated value of 8t . because the maturities are not equally spaced. we fit the yield curve using the three-factor model. Related. and in the bottom panel we plot it by maturity (averaged over time). $2t. 8tJ &8 J We begin by estimating the parameters 2t ' {$1t. The differences are for the most part minor. t i'1 Nt where . $3t} change dramatically at various times. the variation is small relative to the standard error. $2t. This lets us compute the values of the two regressors (factor loadings) and use ordinary least squares to estimate the betas (factors). Both findings suggest that little would be lost by fixing 8t .Fitting Yield Curves As discussed above.11 ˆ ˆ ˆ We will subsequently examine the fitted series {$1t. We verify this claim in Figure 5.it. 2 ˆ 2t ' argmin2 j . Applying Other weightings and loss functions have been explored by Bliss (1997b). and Bates (1999). 8t} by nonlinear least squares.it is the difference at time t between the observed and fitted yields at maturity Ji . that allowing 8t to vary over time can make the estimated {$1t. we found that the sum-ofsquares function is not very sensitive to 8t . The case for fixing 8t becomes very strong when one adds to this the facts that estimation of lambda is fraught with difficulty in terms of getting to the global ˆ ˆ ˆ optimum. as in Nelson and Siegel (1987). In light of the above considerations. Note that. that a slightly better in-sample fit does not necessarily produce better out-of-sample forecasting. Soderlind and Svensson (1997). $3t} in detail. and after some experimentation. We do so for the remainder of this paper. 11 9 . we implicitly weight the most “active” region of the yield curve most heavily when fitting the model. $3t.

which happens occasionally. we attempted to remove illiquid bonds. See.12 The model also has trouble fitting times when the yield curve has multiple interior optima. st) = -0. although it’s not obvious that they would produce better out-ofsample forecasts. slope and curvature. humped. however. slope. as discussed earlier. 15 14 13 12 Forecasting the estimated term structure is done quite commonly. ct) = 0. In Table 3 and Figure 10 we present descriptive statistics for Allowing 8 to vary over time usually improves the fit by only a few basis points. st. $2t. In Figure 6 we plot the implied average fitted yield curve against the average actual yield curve. The residual sample autocorrelations indicate that pricing errors are persistent. if interest centers on forecasting the estimated term structure rather than the actual term structure. complete elimination is not possible. because they persist. or pricing errors.99. The figure confirms our assertion that the three factors in our model correspond to level. lt) = 0. they should vanish from fitted yield changes. As noted in Bliss (1997b). In Figure 7 we dig deeper by plotting the raw yield curve and the three-factor fitted yield curve for some selected dates.97. ct) are the empirical level. it can improve the fit significantly when the short end of the yield curve is steep. Although. downward sloping. Assessing the Fit There are many aspects to a full assessment of the “fit” of our model.ordinary least squares to the yield data for each month gives us a time series of estimates of ˆ ˆ ˆ {$1t. It can be seen that the three-factor model is capable of replicating various yield curve shapes: upward sloping. for example. especially when yields are dispersed. there is a persistent discrepancy between actual bond prices and prices estimated from term structure models. Duffee (2002). and inverted humped. and D($3t . regardless of the term structure estimation method used. Subsequently we will discuss extensions of the three-factor model that would better fit yield curves with multiple interior optima. It does. $2t.13 Overall. the residual plot in Figure 8 indicates a good fit. D($2t . have difficulties at some dates. slope and curvature of the yield curve.14 However. $3t} along with the empirical level. slope and curvature defined earlier. where (lt. 10 . The two agree quite closely. Presumably these discrepancies arise from persistent tax and/or liquidity effects.15 ˆ ˆ ˆ In Figure 9 we plot {$1t. $3t} and a corresponding panel of residuals. The correlations between the estimated factors and the empirical level. then the pricing errors are irrelevant. However. Moreover.99. In Table 2 we present statistics that describe the in-sample fit. however. and ˆ ˆ ˆ curvature are D($1t .

Out-of-Sample Forecasting Performance of the Three-Factor Model A good approximation to yield-curve dynamics should not only fit well in-sample. we can see that the first factor is the most persistent. and that the second factor is more persistent than the third. first as univariate AR(1) processes and second as a VAR(1) process. forecasting improvement from moving to a multivariate model. The AR(1) or VAR(1) models can be viewed as natural benchmarks determined a priori: the simplest great workhorse autoregressive models.16 Finally. 2.t%h/t ' ci % (i $it.8704 at the five percent level. and -2. slope and curvature .55 . but also ˆ ˆ ˆ forecast well out-of-sample.t%h/t % $2. corr($1. The yield forecasts based on underlying univariate AR(1) factor specifications are: 1&e &8J ˆ ˆ yt%h/t(J) ' $1. ˆ ˆˆ The factors are not highly correlated. -2. From the autocorrelations of the three factors. $3)'0.4518 at the one percent level. Because the yield curve depends only on {$1t.t%h/t &e . 8J ˆ ˆ and ci and (i are obtained by regressing $it on an intercept and $i. 1&e &8J &8J ˆ % $3. Similarly. $3t} .the estimated factors. forecasting the yield We use SIC to choose the lags in the augmented Dickey-Fuller unit-root test. Modeling and Forecasting Yield Curve Level. and ˆ ˆ corr($2.5714 at the ten percent level. The autocorrelations are very small. In Figure 11 we provide some evidence on goodness of fit of the models of level. the yield forecasts ˆ ˆ based on an underlying multivariate VAR(1) specification are: 1&e &8J ˆ ˆ yt%h/t(J) ' $1. slope and curvature. 3. Slope and Curvature We model and forecast the Nelson-Siegel factors in two ways.t%h/t ˆ 8J where ˆ $i.07 . $2t.t%h/t &e . ˆ ˆ ˆ i ' 1.t%h/t ˆ 8J 1&e &8J &8J ˆ % $3. the pairwise ˆ ˆ ˆ ˆ correlations between the estimated factors are: corr($1. so that an appropriate multivariate model is close to a stacked set of univariate models. indicating that the models accurately describe the conditional means of level. if any. and that $3 does not. 11 16 . showing residual time series plots and autocorrelation functions. The MacKinnon critical values for rejection of hypothesis of a unit root are -3. Hence we expect little. Augmented Dickey-Fuller ˆ ˆ ˆ tests suggest that $1 and $2 may have a unit roots.t&h .t%h/t % $2. $2)'&0. $3)'&0. 8J ˆ $t%h/t ' c % ' $t.51 .

6 and 12 months. Hence the forecasted yield change is obtained from a regression of historical yield changes on forward premia. $2t.” (2) Slope regression: yt%h/t(J)&yt(J) ' c(J) % ((J) (yt(J)&yt(3)). 60 and 120 months. Let us now describe the competitors in terms of how their forecasts are generated.ˆ ˆ ˆ curve is equivalent to forecasting {$1t. In this section we undertake just such a forecasting exercise. ˆ ˆ ˆ The forecasted yield change is obtained from a regression of historical yield changes on yield curve slopes. (1) Random walk: yt%h/t(J) ' yt(J). 12 17 .17 Note that this is an unrestricted version of the model estimated by Cochrane and Piazzesi. Imposition of the Cochrane-Piazzesi restrictions produced qualitatively identical results. beginning in 1994:1 and extending through 2000:12. using data from 1985:1 to the time that the forecast is made. (4) Cochrane-Piazzesi (2001) forward curve regression: ˆ ˆ ˆ yt%h/t(J)&yt(J) ' c(J) % (0(J) yt(J) % j (k(J) ft (12). the information used to forecast future yields in forward rate regressions is very similar to that in slope regressions. because the forward rate is proportional to the derivative of the discount function. We estimate and forecast recursively. 12. ˆ The forecast is always “no change. for maturities of 3. 36. ˆ k'1 9 12k h h Note that the Fama-Bliss forward regression is a special case of the Cochrane-Piazzesi forward regression. and forecast horizons of h = 1. Note that. ˆ ˆ ˆ where ft (J) is the forward rate contracted at time t for loans from time t%h to time t%h%J . (3) Fama-Bliss forward rate regression: yt%h/t(J)&yt(J) ' c(J) % ((J) (ft (J)&yt(J)). In Tables 4-6 we compare h-month-ahead out-of sample forecasting results from Nelson-Siegel models to those of several natural competitors. $3t} .

although slightly better than the random walk and slope regression forecasts. yt(36)&yt(3). yt(120)&yt&1(120)]) . Our model’s 1-month-ahead forecasting results. yt(120)&yt(3)]). yt(120)&yt(3)]) . ˆ where zt / [yt(3)&yt&1(3). yt(36). In absolute terms. yt(12). yt(60)&yt&1(60). (7) VAR(1) on yield changes: ˆ ˆ zt%h/t ' c % ' zt . including mean. yt(36)&yt(3). yt(120)&yt(3)]) . yt(12)&yt(3). the forecasts appear suboptimal: the forecast errors appear serially correlated. yt(12)&yt&1(12). yt(120)]) . (9) ECM(1) with two common trends: ˆ zt%h/t ' c % ' zt. We define forecast errors at t+h as yt%h(J) & yt%h/t(J) . yt(60)&yt(3). and we report descriptive statistics for the ˆ forecast errors. (8) ECM(1) with one common trend: zt%h/t ' c % ' zt. yt(60)&yt(3). ˆ ˆ ˆ where zt / [yt(3)&yt&1(3). standard deviation. yt(60). are in certain respects humbling. yt(12)&yt&1(12). and autocorrelations at various displacements. ˆ ˆ ˆ (6) VAR(1) on yield levels: ˆ ˆ yt%h/t ' c % ' yt. RMSE comparison at various maturities reveals that our forecasts. 13 .(5) AR(1) on yield levels: yt%h/t(J) ' c(J) % ( yt(J). ˆ ˆ ˆ where zt / [yt(3)&yt&1(3). In relative terms. yt(36)&yt&1(36). (10) ECM(1) with three common trends: zt%h/t ' c % ' zt. are indeed only very slightly better. ˆ where yt / [yt(3). yt(36)&yt&1(36). root mean squared error (RMSE). reported in Table 4. ˆ ˆ where zt / [yt(3)&yt&1(3). yt(60)&yt(3). yt(12)&yt&1(12).

as discussed above. 14 18 . our model’s 12-month ahead forecasts outperform those of all competitors at all maturities. for example. our three-factor model fit to monthly data produces 12month-ahead forecasts that appear much better that those from the Cochrane-Piazzesi model. our 12-monthahead forecasts.18 Finally. because the forecast errors remain serially correlated. In fact. often by a wide margin in both relative and absolute terms. A comparison of our results We report 12-month-ahead forecast error serial correlation coefficients at displacements of 12 and 24 months. are radically improved. pricing errors due to illiquidity may be highly persistent and could be reduced by forecasting the fitted yield curve rather than the “raw” curve. like their 1. Seven of the ten Diebold-Mariano statistics in Table 7 indicate significant 12-month-ahead RMSE superiority of our forecasts at the five percent level. The 1-month-ahead forecast defects likely come from a variety of sources. because the 12-month-ahead errors would naturally have moving-average structure even if the forecasts were fully optimal. often with persistence much stronger than ours. Second. which is appropriately viewed as a victory. instead. we made no attempt to approximate the factor dynamics optimally.and 6-month-ahead counterparts. we note that Duffee (2002) finds that even the simplest random walk forecasts dominate those from the Dai-Singleton (2000) affine model. Second. reported in Table 5. It is worth noting. Our model’s 6-monthahead forecasting results. First. however. Matters improve radically. The strong yield curve forecastability at the 12-month-ahead horizon is. due to the overlap. that related papers such as Bliss (1997b) and de Jong (2000) also find serially correlated forecast errors. are noticeably improved. very attractive from the vantage point of active bond trading and the vantage point of credit portfolio risk management. which therefore appears largely useless for forecasting. or by including variables that may explain mispricing. some of which could be eliminated.Finally. and our model’s 12-month-ahead forecasting results. reported in Table 6. moreover. Hence Duffee proposes a less-restrictive “essentially-affine” model and shows that it forecasts better than the random walk in most cases. Moreover. we simply asserted and fit AR(1) conditional mean models. could be improved upon. in contrast to those at displacements of 1 and 12 months reported for the 1-month-ahead forecast errors. we compare certain aspects of our forecasting results to those in two important and related papers. the Diebold-Mariano (1995) statistics reported in Table 7 indicate universal insignificance of the RMSE differences between our 1-month-ahead forecasts and those from random walks or Fama-Bliss regressions. We find strong forecastability by optimizing the relevant loss function. First. In particular. as the forecast horizon lengthens. we note that we refute the Cochrane-Piazzesi (2001) claim that even an adequate model fit to monthly data would miss the 12-month yield forecastability. for example.

we have: yt ' Z$t % . and we have explored the model’s performance in out-of-sample yield curve forecasting.. H)) E($0 gt ) ' 0 ) ) E($0 0t ) ' 0. which is facilitated by noticing that the model forms a state-space system. and then modeling and forecasting them. slope and curvature. The forecasting results at a 1-month horizon are no better than those of random walk and slope models.t . 15 . but potentially useful and important. diag(Q. $2. Here we discuss several variations and extensions of the basic modeling and forecasting framework developed thus far.. slope and curvature factors. In our view. whereas we forecast the actual yield curve. however. in this paper we followed a two-step procedure. first estimating the level.and Duffee’s. State-Space Representation and One-Step Estimation Let us begin with a technical. it is suboptimal relative to simultaneous estimation.t $t ' c % A$t&1 % R0t 0t . and $3 . reveals that our three-factor model produces larger percentage reductions in outof-sample RMSE relative to the random walk than does Duffee’s best essentially-affine model. Although we believe that we lost little by following the two-step approach.WN(0. t ' 1.. $1. In an obvious vector notation. Concluding Remarks and Directions for Future Research We have provided a new interpretation of the Nelson-Siegel yield curve as a modern three-factor model of level. To maximize clarity and intuitive appeal. extension. all of them represent important directions for future research. T . 4. Our forecasting success is particularly notable in light of the fact that Duffee forecasts only the smoothed yield curve. whereas the results at a 12-month horizon are strikingly superior. .

this risk is directly linked to maturity.t(3) . we can easily generalize the notion of duration to our multi-factor framework. where t < t1 <...t(120) ' 1&e &1208 1&e &1208 &1208 &e 1208 1208 and the transition equation presently features diagonal R and A matrices. 1&e &38 &38 &e 38 1&e &68 &68 &e 68 . Beyond Duration The most important single source of risk associated with holding a government bond is variation in interest rates. slope and curvature factors. longer maturity bonds suffer greater price fluctuations than shorter maturity bonds for a given change in the level of the interest rates.. but also slope and curvature. n.. including allowing for richer dynamics (e. moreover. and exogenous (e. allowing us to produce confidence “tunnels” for the entire yield curve.. macroeconomic) variables.. For a given shift in the yield curve. But in our model. 1 1&e &38 38 1&e &68 68 . Conditional heteroskedasticity in the latent level.. For a coupon bond paying xi units at time ti . the risk of a discount bond with price Pt(J) with respect to the three factors is 16 . j Pt(Ji) xi i'1 It is well-known that duration is a valid measure of price risk only for parallel yield curve shifts. slope and curvature factors. ... which is a weighted average of the maturities of the underlying discount bonds.< tn .. % .. the corresponding risk measure is duration. yt(120) 1 .g. the shifting yield curve. i=1.. may be explicitly accommodated by allowing for a timevarying R matrix. However. Extensions are readily accommodated.In particular. j Ji Pt(Ji) xi i'1 n n i'1 n D ' .g.t(6) .. . non-diagonal A and R matrices).. yield curves typically shift in non-parallel ways involving not only level. with price given by Pct ' j Pt(Ji) xi with Ji ' ti&t as required to eliminate arbitrage opportunities. 2. For a discount bond. Maximum likelihood estimates are readily obtained via the Kalman filter in conjunction with the prediction-error decomposition of the likelihood. and certainly in the real world. that is.. as are optimal extractions and forecasts of the latent level. $1t $2t $3t ... the measurement equation is 1 yt(3) yt(6) .

generalized duration. is an obvious way to take a position reflecting views regarding expected shifts in the level of the yield curve. as with any risk management tool. slope and curvature shifts. involving increasing portfolio duration when rates are expected to decline. The traditional duration measure corresponds to maturity. A bullet portfolio has maturities centered at a single point on the yield curve. by splitting bond price risk into components associated with yield curve level. where wi ' . but it nevertheless often remains straightforward to take appropriate positions reflecting views on likely movements. For a coupon bond paying xi units at time ti .. corresponding to slope and curvature risk. associated with the three loadings on d$1t . for example. Given that our three-factor model provides an accurate summary of yield-curve Pct dynamics. Active bond portfolio managers attempt to profit from their views on changes in the level and shape of the yield curve.< tn .& d Pt(J) Pt(J) ' J dyt(J) ' J d$1t % 1&e &8J d$2t % 8 1&e &8J & Je &8J d$3t . The so-called interest rate anticipation strategy. Ingersoll and Ross (1979). d$2t and d$3t in the above equation. generalized duration may also be used in a speculative mode. that is why it is an adequate risk measure only for parallel yield curve shifts. In future work. where t< t1 <. and conversely. In particular. J.. and a barbell 17 Pt(Ji) . now feature prominently as well. suggests using separate forecasts of those components to guide active trading strategies. with immediate application to fixed income risk management. 8 Hence there are now three components of price risk. . the commonly observed cases of (1) a downward shift combined with a steepening. Consider. our generalized duration should provide an accurate summary of the risk exposure of a bond portfolio. 2. Hence. i=1. The notion of generalized duration is readily extended to coupon bonds. n. which is the loading on the level shock d$1t . comparing generalized duration to the traditional Macaulay duration and to the stochastic duration of Cox. for example with bullet or barbell portfolios. Active Bond Portfolio Management Hedging and speculation are opposite sides of the same coin. The yield curve can of course shift in ways more subtle than simple level shifts.. Generalized duration of course still tracks “level risk. and (2) an upward shift combined with a flattening..” but two additional terms. we define the corresponding three-factor generalized duration as & d Pct Pct ' j wixiJi dyt(Ji) ' j wixiJi d$1t % j wixi i'1 i'1 i'1 n n n 1&e 8 &8Ji d$2t % j wixi i'1 n 1&e 8 &8Ji & w ix iJ ie &8Ji d$3t.. we plan to pursue this idea.

For example. as well as the interaction between different countries’ term structures. for recent developments see Estrella and Mishkin (1998) and Stock and Watson (2000). the bullet will outperform if the yield curve steepens with long rates rising relative to short rates. At issue. considering that the long rate is approximately the average of expected future short rates plus a term premium. Standard central bank reaction functions suggest that short rates should be related to output and inflation. $3t} and various macroeconomic fundamentals. Perhaps the reason is the restrictions placed on term premia in affine models. but also level and curvature. it will be of interest to explore the interaction between the term structure and various financial. the Taylor (1993) rule is typically written as ( ¯ i t ' c % NB (Bt&B() % Nz (zt&zt ) ¯ it ' 2 it&1 % (1&2) i t . B( is the target inflation rate. Now consider the effects of the term structure on macroeconomic fundamentals. it is the actual short rate set by the policy maker. as opposed to macroeconomic. (2001) explore the predictive content of forward premia for 18 ( . that not only slope. in both directions. Finally. is how macroeconomic fundamentals affect yields as one moves out the yield curve. ¯ where i t is the target short interest rate. Consider first the effects of macroeconomic fundamentals on the term structure. in general. Clarida et al. It will be interesting. the barbell will almost surely outperform because of the positive effect of capital gains on long term bonds. $2t. Conversely. The predictive content of the yield curve has long been recognized. because of the capital loss on the longer term bonds in the barbell portfolio. Typically only the term structure slope is used as a leading indicator. Our analysis suggests. if the yield curve flattens with long rates falling relative to short rates. Along these lines. Links of the Yield Curve to Macroeconomic and Other Financial Indicators It will be of interest to explore the links between various yields and macroeconomic fundamentals. within the framework developed here. Bt is the inflation rate. This result is intriguing.portfolio has maturities concentrated at two extreme points on the yield curve. are necessary for a full summarization of the information in the term structure. therefore. For example. to see whether variables such as curvature have marginal predictive content. variables. zt is output. with one maturity shorter and the other longer than that of the bullet portfolio. and zt is target output. This effectively amounts to ˆ ˆ ˆ characterizing the relationships among {$1t. in contrast. Hence. however. Ang and Piazessi (2000) find in an affine environment that macroeconomic fundamentals move yields at the short but not the long end of the curve.

They begin with the instantaneous forward rate curve. the forward premium is just the interest rate differential. Bliss (1997b) extends the model to include two decay parameters. By “inconsistent” we mean that if we start with a forward rate curve that satisfies Nelson-Siegel. 82tJ &8 J The Soderlind-Svensson model allows for up to two humps in the yield curve. Hence the ultimate issue is whether the term structure of credit spreads can be distilled efficiently via the three-factor model advocated in this paper. 82tJ &8 J Obviously the Bliss curve collapses to the original Nelson-Siegel curve when 81t'82t . Soderlind and Svensson (1997) also extend Nelson-Siegel to allow for two decay parameters. Björk and Christensen (1999) show that in the Heath-Jarrow-Morton (1992) framework with deterministic volatility. For example. whereas the original Nelson-Siegel model allows only one. ft(J) ' $1t % $2t e which implies the yield curve yt(J) ' $1t % $2t 1&e 1t 81tJ &8 J &81tJ % $3t 81t Je &81tJ % $4t 82t Je &82tJ . Generalizations to Enhance Flexibility and to Maintain Consistency with Standard Interest Rate Processes A number of authors have proposed extensions to Nelson-Siegel that enhance flexibility. albeit in a different way. ft(J) ' $1t % $2t e we obtain the yield curve yt(J) ' $1t % $2t 1&e 1t 81tJ &8 J &81tJ % $3t 82t Je &82tJ . Nelson-Siegel forward-rate dynamics are inconsistent with standard interest rate processes. % $3t &8 J 1&e 1t &e 1t 81tJ &8 J % $4t &8 J 1&e 2t &e 2t . Beginning with the instantaneous forward rate curve. % $3t &8 J 1&e 2t &e 2t . and if interest rates subsequently evolve according to the Ho-Lee or Hull19 . But by covered interest parity. A number of authors have also considered generalizations of Nelson-Siegel to maintain consistency with arbitrage-free pricing for certain short-rate processes. such as those of Ho and Lee (1986) and Hull and White (1990).subsequent spot exchange rate movements. and then used to enhance spot exchange rate forecasts. and of course there are many interest rate differentials as one moves out the term structure.

19 Björk (2000). there is no guarantee of better outof-sample forecasting performance. but also promote consistency with interest rate dynamics associated with modern arbitrage-free models. but also with the two-factor models studied in Heath. ft(J) ' $1t % $2t J % $3t e &8J % $4t Je &8J % $5t e &28J. Björk and Christensen (1999). and if doing so would not only improve forecasting performance. is consistent not only with Ho-Lee and Hull-White. 2000) extends this negative result to stochastic volatility environments. Filipovic (1999. For example. Nevertheless. Note that 8 is constant in the above expression. although the Bliss and Soderlind-Svensson extensions can have in-sample fit no worse than that of Nelson-Siegel. both the parsimony principle and accumulated experience suggest that parsimonious models are often more successful for out-of-sample forecasting. the desirability of the above generalizations of Nelson-Siegel is not obvious.White models. the serially correlated and forecastable errors produced by our three-factor model reveal the potential for improvement. providing further justification for our assumption of constancy in our earlier empirical work. Jarrow and Morton (1992) under deterministic volatility. perhaps by adding additional factors. 20 . which is why we did not pursue them here. 20 19 See Diebold (2001). however. From the perspective of interest rate forecasting accuracy. show that a five-factor variant of the Nelson-Siegel forward rate curve. so much the better. however. although consistency with some historically-popular interest rate processes is perhaps attractive ceterus paribus.20 Similarly. We look forward to exploring this possibility in future work. Indeed. then the corresponding forward curves will not satisfy Nelson-Siegel. because they include Nelson-Siegel as a special case. Björk and Landén (2000) and Björk and Svensson (2001) and provide additional insight into the sorts of term structure dynamics that are consistent with various forward curves. it is not clear that insisting upon it would improve forecasts.

” Manuscript. L. A. (1953). P. D. “A Geometric View of Interest Rate Theory. 82. and Piazzesi. University of Pennsylvania. T. “The Out-of-Sample Success of Term Structure Models as Exchange Rate Predictors: A Step Beyond. Björk. Foresi. 52. Björk. 51-61. 97–231. (1997). (1999). C. 51.W. M. and Ross. and Piazzesi. “Interest Rate Dynamics and Consistent Forward Rate Curves. (2000). “Time-Consistent No-Arbitrage Models of the Term Structure. Stockholm School of Economics. Department of Finance. Bates. Brandt. London: Blackwell Publishers. “A Simple Approach to Three Factor Affine Term Structure Models. Björk. A. Kellogg School. 11. R. Cox. Balduzzi. Ingersoll.” Mathematical Finance. Bliss.G. L. (1997a).. “Bond Risk Premia. 9. D. “Stochastic Volatility and Mean Drift in the Short Term Interest Rate Diffusion: Source of Steepness. S. and Lund. T.” Working Paper 214.” Advances in Futures and Options Research.E. Sarno. E. (1979).” Journal of Business.R. “Duration and Measurement of Basis Risk. Chen. and Valente. T. forthcoming.References Andersen.” Journal of Fixed Income. G.” Manuscript. (1996). M.” Manuscript.” Economic Review. Taylor.. Stochastic Mean and Stochastic Volatility . (1997b). R. (2000). (2000).A. University of Chicago and UCLA. Columbia University and University of Warwick.” Working Paper.P. S. 16-33.” Manuscript. R. and Hilbert.” Handbook of Mathematical Finance. (2001). J. 9. (2001). S.H. “Movements in the Term Structure of Interest Rates. L. Level and Curvature in the Yield Curve. Das. Columbia University. (2001). 205-243. “On the Existence of Finite Dimensional Realizations for Nonlinear Forward Rate Models. . “A No-Arbitrage Vector Autoregression of Term Structure Dynamics with Macroeconomic and Latent Variables.. T.H. Cochrane. Björk. R.” Mathematical Finance. 6. New York: John Wiley. M. Cambridge: Cambridge University Press. Federal Reserve Bank of Atlanta. T. Clarida.” Carnegie-Rochester Conference Series on Public Policy.. Courant. M. 43-53. and Yaron. Northwestern University. Ang. Bliss. “On the Construction of Finite Dimensional Realizations for Nonlinear Forward Rate Models. 323-348. Methods of Mathematical Physics... and Landén. “Financial Markets’ Assessment of EMU. and Svensson. J. J. (2001).A Three Factor Model of the Term Structure of Interest Rates and its Application to the Pricing of Interest Rate Derivatives. (1996). and Sundaram. B.C. “Testing Term Structure Estimation Methods. and Christensen. 229-269. J. (1999).

J. 1-27. Q. (1998).” American Economic Review. and Sharpe. F. (2000). 45-61.” Journal of Business and Economic Statistics. “Evaluating Density Forecasts. “A Note on the Nelson-Siegel Family.” Review of Economics and Statistics. 57. S. J. Hahn. Diebold. J.S. Filipovic. 379-406. (1999).E. and Ross. Diebold. Duffie. T.” Journal of Business and Economic Statistics. A. “A Theory of the Term Structure of Interest Rates. (1998). 680-692. P.. Diebold. F. F.” Mathematical Finance.” Mathematical Finance.A. Dai.S. “Idiosyncratic Variation of Treasury Bill Yields. 131-157. and Lown. A. “Post-Deregulation Bank Deposit Rate Pricing: The Multivariate Dynamics. Duffee. 527-551. A. 77. (2000).” International Economic Review. 281-293. 405-443. E. F. 109. 55. 13.X. D. Duffee. Diebold. Recessions: Financial Variables as Leading Indicators. Cincinnati: South-Western. Estrella. and Santa-Clara. “The Information in Long-Maturity Forward Rates. (2002).” Journal of Finance. (2000). R. (1999). 1943-1978.X. “Time Series and Cross Section Information in Affine Term Structure Models. 18. and Mariano. “An Indicator of Future Inflation Extracted from the Steepness of the Interest Rate Yield Curve along its Entire Length. 80.S.S. and Kan. and Tay. F.. 9. 6. (1990). S. Elements of Forecasting (Second Edition). 385-407.” Journal of Finance. (1985). and Mishkin. (2001). and Bliss. 53.” Quarterly Journal of Economics. “Multivariate Density Forecast Evaluation and Calibration in Financial Risk Management: High-Frequency Returns on Foreign Exchange.S. “Predicting U. F.X. 517-530 . R.” Review of Economics and Statistics. de Jong. Ingersoll. Frankel.” Journal of Finance. (1995). A. 253-263. 661-673.S. K. “Term premia and interest rate forecasts in affine models. and Tay. 31.” Econometrica.X.” Journal of Financial and Quantitative Analysis. 81.Cox. Fama. Filipovic. 300-314. “Specification Analysis of Affine Term Structure Models. “Exponential-Polynomial Families and the Term Structure of Interest Rates. Gunther. 349-359. G. F. 6. C. 863-883. 8. R.. Diebold.” Bernoulli. 51. and Singleton. (1994). “The Dynamics of the Forward Interest Rate Curve: A Formulation with State Variables. (1996). G. “A Yield-Factor Model of Interest Rates. D.X.C. (1987).” Journal of Business and Economic Statistics. D. “Comparing Predictive Accuracy. F. de Jong. with Applications to Financial Risk Management. (1996). J. (1999). 39.

C. D. 265-275.H.D. (1986). J.-B. 5. “Bond Pricing and the Term Structure of Interest Rates: A New Methodology for Contingent Claims Valuation. “Long-term Behavior of Yield Curves.” Journal of Monetary Economics. 473-489. 37. C. H. and Morton. C. and Granger.” Manuscript. “Forecasting Output and Inflation: The Role of Asset Prices.H. 39.R.. “Term Structure Movements and the Pricing of Interest Rate Contingent Claims. T.S. McCulloch. A. Litzenberger. G. 1947-1991. J. (1988). 105-110. S. Siegel.R Rao and G. and White. and Weir. 74.G. Pagan. .” Carnegie-Rochester Conference Series on Public Policy. Taylor.S. Squassi. 347-366. Handbook of Statistics. A.R. and Kwon. (1996). (1982).J.F. (1977). “A Model-Selection Approach to Assessing the Information in the Term Structure Using Linear Models and Artificial Neural Networks. Maddala (eds.A. (1990). 77-105.” Review of Financial Studies. N. “Spline Models of the Term Structure of Interest Rates and Their Applications. McCulloch.S. Goldman.E. (1995). (1992). J. Amsterdam: North-Holland. Ohio State University.. and Martin. “The Tax Adjusted Yield Curve. A.D.. Anderson. “Term Structure Modeling Using Exponential Splines.. Heath. J.” Working Paper. 3. A.R. Hall.M.W. 383-430. Term Structure Data. Soderlind.R.” in C.). O. “An Equilibrium Characterization of the Term Structure. “Pricing Interest-Rate-Derivative Securities.” Journal of Finance.” Journal of Business and Economic Statistics. 339-348.” Econometrica.S. Harvard University and Princeton University. 13. 195-214 Vasicek. 40. and Svensson. and Fong. (1995).” Journal of Financial Economics. 116-126. (1975). 60. Stock. (1997). 573-592. G.W. 41. Vasicek. J.” Journal of Business. and Lee. (1992). O.” Journal of Financial and Quantitative Analysis. A. (1993). H.O. A. Shea. (1987). and White. “U.” Working Paper 93-6. 91-118. “Parsimonious Modeling of Yield Curves. P. V. and Nelson. L.” Journal of Business and Economic Statistics. 60. H. Sachs and Company. “New Techniques to Extract Market Expectations from Financial Instruments. “Discretion versus Policy Rules in Practice. N. “A Cointegration Analysis of Treasury Bill Yields. M. Hull. R.” Journal of Finance.” Journal of Finance. A. H. Nelson. Jarrow. 10. “Benchmarking the Expectations Hypothesis of the Interest-Rate Term Structure: An Analysis of Cointegration Vectors. 811-830.B (1993). R. and Siegel.” Review of Economics and Statistics. (2000). and Watson. 23.H. (1992).Hall. “Modeling the Term Structure. Ho. 1011-1029. 30. Swanson.F. 177-188.

352 4.663 4.439 1.482 3.448 0.131 9.496 1.504 1.896 D(12) ˆ 0.569 0.891 2.288 3.960 0.042 -0.488 1.968 0.513 0.953 0.432 1.953 0.462 1.957 0.274 10.428 1.965 0.502 0.142 7.049 -0.060 0.748 10.653 11.115 0.413 10.457 1.015 .313 11.976 0.978 0.956 0.732 2.963 0.060 1.308 6.664 11.426 0.984 3.973 0.308 4.343 9.961 0.226 0.951 0.429 4.785 5.501 1.479 0.555 0.484 1.948 0. Yield Curves Maturity (Months) 3 6 9 12 15 18 21 24 30 36 48 60 72 84 96 108 120 (level) slope curvature Mean 5.225 6.337 D(30) ˆ -0.928 7.405 0.430 1.067 6.269 11.838 6.133 0.270 7.443 -0.375 6.550 6.467 0.188 10.428 -0.457 0.089 0.005 0.787 11.254 1.043 4.777 4.021 0.471 0.372 0.384 4.837 Maximum 9.439 1.638 3.907 6.683 9.204 4.324 9.602 D(1) ˆ 0.Table 1 Descriptive Statistics.527 0.475 0.954 0.107 3.545 0.624 -0.648 Minimum 2.752 -1.468 0.492 1.081 Standard Deviation 1.294 0.539 0.464 0.482 1.951 0.226 7.082 7.988 10.464 1.841 11.401 6.425 1.417 0.512 11.336 0.413 1.079 -0.347 4.481 0.213 0.433 4.190 0.391 0.630 5.969 0.644 6.953 0.454 0.

199 -0.777 0.056 -0.044 0.080 0.597 0.218 0.065 0.161 -0.243 0.057 0.052 0.754 0.758 0.183 -0.398 0.156 0.070 . Max.062 0.066 0.055 0. Yield Curve Residuals Maturity (Months) 3 6 9 12 15 18 21 24 30 36 48 60 72 84 96 108 120 Mean Standard Deviation 0.163 0.067 0.052 0.063 0.063 -0.032 0.032 0.139 0.111 0.046 -0.200 -0.019 0.120 0.059 0.027 -0.052 0.029 0.001 0.157 0.062 0.059 0.399 0.218 0.263 0.101 0.202 -0.048 -0.139 -0.044 0.266 -0.098 0.010 0.045 0.204 -0.697 0.332 -0.080 0.278 0.190 -0.045 0.042 0.052 0.030 0.050 0.071 Min.013 0.062 0.061 0.080 0.203 -0.011 0.Table 2 Descriptive Statistics.018 -0.056 0.045 0.360 -0.079 0.053 0.044 0.200 -0.037 0.048 0.669 0.067 0.650 0.026 -0.186 0.322 0.033 -0. MAE RMSE D(1) ˆ D(12) ˆ D(30) ˆ -0.496 0.091 -0.370 0.073 0.904 0.252 -0.035 0.175 -0.053 0.037 -0.046 0.082 0.176 -0.128 0.144 -0.216 0.021 0.017 -0.016 -0.230 0.081 0.623 0.070 -0.667 0.057 0.132 0.033 0.047 0.082 0.563 0.257 0.141 -0.040 0.018 -0.064 0.133 -0.058 -0.036 0.058 0.013 -0.164 0.321 -0.589 0.247 -0.067 0.000 -0.041 0.026 0.704 0.160 -0.212 0.081 0.072 0.291 0.046 0.259 -0.256 0.239 -0.267 0.081 0.047 0.020 -0.080 0.165 0.

234 D(1) ˆ 0.006 ADF -2.919 4.454 -0. Estimated Factors Factor ˆ $1t ˆ $2t ˆ $3t Mean 7.957 0.524 1.511 0. Dev. 1.249 Maximum 12.Table 3 Descriptive Statistics.410 -1.098 -0.969 0.205 -3.687 Minimum 4.901 D(12) ˆ 0.452 0.608 1.082 -0.162 Std.353 D(30) ˆ 0.427 -5.579 -2.088 0.516 .616 -5.

277 0.024 0.153 -0.425 0.286 0.318 0.260 D(1) ˆ NA 0.215 D(12) ˆ 0.021 0.062 Std.328 0.373 0.179 0.213 -0.341 0.233 0. 0.313 0.340 0.053 -0.159 0.288 0.260 D(1) ˆ 0.041 Std.056 -0.242 0.240 0.259 D(12) ˆ 0.276 0.254 D(1) ˆ 0.131 -0.287 0.013 Std.066 0.145 -0.242 0.273 0.276 0.254 RMSE 0.176 0.176 0.279 0.159 . 0.032 0.033 0.048 0. Dev.251 RMSE 0.279 0.150 -0.159 Fama-Bliss Forward Rate Regression Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.273 0.Table 4 Out-of-Sample 1-Month-Ahead Forecasting Results Nelson-Siegel with AR(1) Factor Dynamics Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean -0.148 -0.146 -0.286 0.007 -0.038 0. Dev.017 -0.260 RMSE NA 0.279 0.045 0.036 -0.284 0.133 -0. Dev.145 RMSE 0.170 0.172 0.200 D(12) ˆ 0.019 0.157 -0.252 Random Walk Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.245 D(12) ˆ NA -0. 0.275 0.280 0.125 -0.117 -0.235 0.292 0.333 0.011 Std.247 0.220 0.115 Slope Regression Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean NA 0.247 0.066 0.116 -0.254 D(1) ˆ 0.091 -0.285 0.332 0.003 -0.023 -0. NA 0. Dev.380 0.178 0.241 0.277 0.236 0.

162 -0.029 0.429 0.287 0.262 0.176 0.030 .Table 4 (Continued) Out-of-Sample 1-Month-Ahead Forecasting Results Cochrane-Piazzesi Forward Curve Regression Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean NA -0.230 0.274 0.281 D(1) ˆ NA 0.289 0.271 D(12) ˆ NA -0.060 -0.232 0.264 D(1) ˆ -0.021 0. 0.128 -0.292 0.084 -0.177 0.310 0.277 0.377 0.305 0.017 D(12) ˆ 0.025 -0.026 -0.154 -0. Dev.149 -0.030 -0.077 0.276 0.437 0.013 -0.038 -0.277 0.447 0.263 RMSE 0.019 0.238 0.239 0.042 0.274 RMSE 0.364 0.125 -0.176 0.034 -0.258 D(1) ˆ 0.041 -0.043 0. Dev.090 Std.224 D(12) ˆ 0.144 VAR(1) on Yield Levels Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean -0.303 0.282 0.310 D(12) ˆ 0. 0.263 0.123 VAR(1) on Yield Changes Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean 0.257 RMSE NA 0.064 -0.181 0.049 -0. NA 0.054 Std.300 0.182 0.276 0.088 -0.345 0.020 Std.133 -0.341 0.288 D(1) ˆ 0. 0.276 0.068 -0.276 0. Dev.127 -0.276 0.147 -0.241 0.002 -0.108 -0.113 Std.238 0.176 0.302 0.280 0.229 0.010 -0.157 0.229 0. Dev.097 Univariate AR(1)s on Yield Levels Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.252 RMSE 0.026 0.005 -0.156 -0.

517 0.879 0.235 -0.209 0.861 0.008 0.048 -0.072 -0.758 0.255 0. Dev.220 0.223 Std.739 0.198 -0. 0.944 0.494 0.181 0.038 D(18) ˆ -0.096 D(18) ˆ NA -0.997 0.256 Fama-Bliss Forward Rate Regression Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.168 0.932 0.373 0.150 -0.964 0.422 0.825 D(6) ˆ 0.251 Std.Table 5 Out-of-Sample 6-month-Ahead Forecasting Results Nelson-Siegel with AR(1) Factor Dynamics Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.301 0.758 RMSE 0.116 0.103 0.381 0.190 -0.985 0.052 -0. NA 0.249 -0.758 D(6) ˆ 0.510 0. 0.758 0.845 D(6) ˆ NA 0.131 -0.211 -0.018 0.939 0.092 0.069 0.564 0.750 0.174 -0.962 0.832 RMSE NA 0.145 Std.271 RMSE 0.189 -0.099 0.669 0.113 -0.211 -0.050 0.020 Std.821 0. Dev.902 0.268 .958 0.721 D(6) ˆ 0.248 -0.109 0.049 0.083 0.777 0.208 0.860 0.058 D(18) ˆ -0.676 Random Walk Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.873 0. Dev.811 0.549 0.150 -0.288 Slope Regression Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean NA 0.748 0.779 0.220 0.794 RMSE 0. 0.019 D(18) ˆ -0.214 -0.605 0.173 -0. Dev.139 0.914 0.194 0.273 -0.656 0.281 0.

240 -0.110 -0.926 0. 0.169 -0.276 0.181 -0.286 Std.800 0.210 -0. 0.673 0.659 0.934 0.059 0.246 0.312 0.789 0.837 RMSE NA 0.192 D(18) ˆ NA -0.951 0. Dev.845 0. 0.731 0.048 0.211 -0.218 -0.160 -0.917 0.073 D(18) ˆ -0.955 0.193 0.345 Std.253 -0.725 0.061 0.981 0. Dev.919 0.809 RMSE 0.941 0.210 -0.905 D(6) ˆ NA 0. NA 0.755 D(6) ˆ 0.310 0.405 0.287 Univariate AR(1)s on Yield Levels Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.179 0.661 0.259 .289 0.220 0.811 RMSE 0.907 D(6) ˆ 0.043 D(18) ˆ -0.845 0.195 -0.293 VAR(1) on Yield Changes Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean 0.801 0.144 -0.910 0.406 Std.831 D(6) ˆ 0.218 -0.258 -0.901 0.273 -0.210 -0.859 0.193 0.256 -0. Dev.802 0.319 0.242 -0.075 0.949 0.584 0.937 0.224 0.155 -0.880 0.140 0.900 0.707 0.270 -0.133 0. Dev.910 0.539 0.172 0.122 0.192 Std.030 -0.224 -0.699 RMSE 0.138 -0.137 D(18) ˆ -0.278 VAR(1) on Yield Levels Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean -0.160 -0.155 -0.Table 5 (Continued) Out-of-Sample 6-month-Ahead Forecasting Results Cochrane-Piazzesi Forward Curve Regression Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean NA -0.

337 -0.261 D(12) ˆ NA -0.163 -0.393 1.235 1.035 0.403 1.978 0.433 D(24) ˆ 0.508 D(24) ˆ -0.000 1.187 -0.412 -0.039 -0.130 -0.042 . NA 1.416 0.419 -0.739 0.316 1.102 1.255 -0.024 0.111 -0.896 0.210 1.065 -0.930 1.033 Std.268 -0.037 Slope Regression Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean NA 0.531 Std. 0.841 0.003 -0.673 -0.069 RMSE 1.362 Std.191 1.332 -0.004 0.268 D(24) ˆ NA -0.288 -0.910 0.072 0.237 1.388 0.184 1.481 -0. Dev.293 1.036 -0.214 1.464 1. Dev.724 0.224 1.823 0.305 1.299 D(24) ˆ -0.385 1.825 RMSE 0.268 1.641 0.197 1.Table 6 Out-of-Sample 12-month-Ahead Forecasting Results Nelson-Siegel with AR(1) Factor Dynamics Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.317 -0.150 0.132 1. 0.056 Random Walk Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.408 -0.109 -0.315 -0.173 -0.236 0.265 -0.515 0.918 0.019 0.526 1.051 RMSE 1.118 -0.052 D(12) ˆ -0.003 Nelson-Siegel with VAR(1) Factor Dynamics Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean -0.208 RMSE NA 1.015 0.416 -0. 1.024 0. Dev.123 -0.918 0.019 1.060 0.212 -0.721 Std.001 -0. Dev.981 D(12) ˆ -0.279 D(12) ˆ -0.463 -0.576 -0.

808 0.760 Std.325 1.028 -0.567 1.043 1.352 -0.096 -0.007 1.198 RMSE 1.039 0.307 D(24) ˆ NA -0. Dev.004 0.030 D(12) ˆ -0.021 -0.006 1.961 0.020 VAR(1) on Yield Levels Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean -0.046 -0.219 -0.875 0.062 Cochrane-Piazzesi Forward Curve Regression Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean NA -0.Table 6 (Continued) Out-of-Sample 12-month-Ahead Forecasting Results Fama-Bliss Forward Rate Regression Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean 0.142 -0.060 RMSE 1.099 -0.529 -0.010 1.266 1.213 -0.291 -0.010 . Dev.058 -0.547 1. NA 1.378 1.035 0.996 0.058 0.835 RMSE 0.970 1.845 0.204 1.274 -0. 1.301 -0.381 1.271 -0.587 0.328 D(24) ˆ -0.334 1.301 -0.363 1.942 0.322 -0.345 -0.317 1. 0.162 -0.079 -0.348 -0. Dev.431 D(24) ˆ -0.327 D(12) ˆ NA -0.040 0.002 1.073 -0.603 Std.317 D(12) ˆ -0.276 -0.061 0.020 Univariate AR(1)s on Yield Levels Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.433 -0.285 1.275 1.467 -0.275 1.012 -0.182 -0.540 -0.953 0.240 1.113 -0.295 D(12) ˆ -0.403 D(24) ˆ -0.744 Std.246 0.276 1.390 -0.201 1.380 -0.225 1. 1.547 Std.015 -0.484 1.377 -0.179 -0.088 RMSE NA 1.330 1. Dev.746 0.

261 1.246 -0.250 1. 1.082 -0.248 -0.274 D(24) ˆ -0.014 0.207 D(12) ˆ -0.035 RMSE 1.169 Std.009 ECM(1) with Two Common Trends Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.072 ECM(1) with one Common Trend Maturity (J) 3 months 1 year 3 years 5 years 10 years Mean 0.425 0.426 1.519 1.069 -0.191 1.408 Std.127 -0.577 1.034 -0. Dev.013 -0.288 -0.486 1.290 1.278 -0.286 -0.064 0.186 1. Dev.228 1.191 D(24) ˆ -0.051 0.717 0.613 0.136 RMSE 1.323 1.546 0.072 -0.341 1.163 -0.459 -0.982 1.281 1.559 0.155 1.203 1. 1.379 0.070 0.304 -0.245 -0.491 D(24) ˆ -0. 0.278 -0.240 1.247 1.231 1.868 0.223 -0.778 0.224 D(24) ˆ -0. Dev.123 -0.786 0.175 -0.810 0. 0.220 Std.095 RMSE 1.306 0.704 0.037 1.003 0.376 1.143 1.086 .738 0.321 1. Dev.453 1.296 1.Table 6 (Continued) Out-of-Sample 12-month-Ahead Forecasting Results VAR(1) on Yield Changes Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean 0.398 1.141 RMSE 1.289 -0.108 D(12) ˆ -0.063 Std.058 D(12) ˆ -0.072 1.249 1.273 1.239 -0.480 1.951 1.041 0.015 ECM(3) with Three Common Trends Maturity ( J ) 3 months 1 year 3 years 5 years 10 years Mean 0.399 -0.767 0.627 0.033 -0.129 -0.236 1.068 -0.143 D(12) ˆ -0.586 0.028 -0.

04* -2.27 -0.18* -1.18 -0.90* -1.61 -0.65* -2.56 -0.58 0.63 against FB -2.34 12-Month Horizon against RW -1.35 3 months 1 year 3 years 5 years 10 years .43* -2.57 0.97 0.64 -0.Table 7 Out-of-Sample Forecast Accuracy Comparisons 1-Month Horizon Maturity (J) against RW -0.11* -1.02 0.31* -2.49 against FB 0.

12. Notes to Table 5: We present the results of out-of-sample 6-month-ahead forecasting using eight models. Negative values indicate superiority of our three-factor model forecasts. Notes to Table 6: We present the results of out-of-sample 12-month-ahead forecasting using twelve models. $2t . and we report the mean. Notes to Table 3: We fit the three-factor Nelson-Siegel model using monthly yield data 1985:01-2000:12. and the curvature as the twice the 2-year yield minus the sum of the 3month and 10-year yields. and the three columns to its left contain sample autocorrelations at displacements of 1. 12. The last three columns contain residual sample autocorrelations at displacements of 1. 8tJ &8 J using monthly yield data 1985:01-2000:12. Notes to Table 4: We present the results of out-of-sample 1-month-ahead forecasting using eight models.0609. and we present descriptive statistics for the three estimated factors $1t . the slope as the difference between the 10-year and 3-month yields. beginning in 1994:1 and extending through 2000:12. We define forecast errors at t +1 as ˆ yt%1(J) & yt%1/t(J) . slope and curvature. The sample period is 1985:01-2000:12. 12. We estimate all models recursively from 1985:1 to the time that the forecast is made. Notes to Table 2: We fit the three-factor model. and 30 months. standard deviation and root mean squared errors of the forecast errors. The last three columns contain sample autocorrelations at displacements of 1. We estimate all models recursively from 1985:1 to the time that the forecast is made. The null hypothesis is that the two forecasts have the same mean squared error. . and we present descriptive statistics for the corresponding residuals at various maturities. where we define the level as the 10-year yield. yt(J) ' $1t % $2t 1&e 8tJ &8tJ % $3t &8 J 1&e t &e t . and asterisks denote significance relative to the asymptotic null distribution at the ten percent level. The last column contains augmented Dickey-Fuller (ADF) unit root test statistics. as well as their sixth and eighteenth sample autocorrelation coefficients. and we report the mean. We define forecast errors at t +12 as ˆ yt%12(J) & yt%12/t(J) . and we report the mean. Notes to Table 7: We present Diebold-Mariano forecast accuracy comparison tests of our three-factor model forecasts (using univariate AR(1) factor dynamics) against those of the Random Walk model (RW) and the FamaBliss forward rate regression model (FB). and 30 months. We estimate all models recursively from 1985:1 to the time that the forecast is made. standard deviation and root mean squared errors of the forecast errors. beginning in 1994:1 and extending through 2000:12. with 8t fixed at 0. as well as their first and twelfth sample autocorrelation coefficients.0609. and 30 months. as described in detail in the text. We define forecast errors at t +6 as ˆ yt%6(J) & yt%6/t(J) . with 8t fixed at ˆ ˆ ˆ 0. and for the yield curve level. as described in detail in the text. as described in detail in the text. and $3t . standard deviation and root mean squared errors of the forecast errors.Notes to Tables Notes to Table 1: We present descriptive statistics for monthly yields at different maturities. as well as their twelfth and twenty-fourth sample autocorrelation coefficients. beginning in 1994:1 and extending through 2000:12.

6 β 2 Loadings 0.0609. and $3t . $2t .4 0. 8tJ &8 J . and J denotes maturity. yt(J) ' $1t % $2t 1&e 8tJ &8tJ % $3t &8 J 1&e t &e t 8tJ &8 J where the three factors are $1t . 1&e 8tJ &8tJ .8 Loadings 0. the associated loadings are 1. We fix 8t = 0. in Months) 80 100 120 Notes to Figure 1: We plot the factor loadings in the three-factor model. and &8 J 1&e t &e t .Figure 1 Factor Loadings 1 β 1 Loadings 0.2 β 3 Loadings 0 0 20 40 60 τ (Maturity.

84. 9. 30. .Figure 2 Yield Curves.01-2000. 1985. 96. 48. 21. 24. and 120 months. 108.12 Notes to Figure 2: The sample consists of monthly yield data from January 1985 to December 2000 at maturities of 3. 60. 6. 15. 18. 36. 12. 72.

4 0.0 -0. We define the level as the 10-year yield.6 0.8 0.6 0.0 -0.0 -0. and in the right panel we plot their sample autocorrelations.2 0.2 -0.0 0.4 0.0 0.Figure 3 Yield Curve Level.0 0. to a displacement of 60 months. slope and curvature.2 -0.8 0.2 -0. along with Bartlett’s approximate 95% confidence bands. .4 Slope Autocorrelation 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Time Curvature Plot 12 10 Curvature (Percent) 8 Autocorrelation 6 4 2 0 -2 86 88 90 92 94 96 98 00 1.2 0.6 0. the slope as the difference between the 10-year and 3-month yields.4 Curvature Autocorrelation 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Time Notes to Figure 3: In the left panel we present time-series plots of yield curve level. and the curvature as the twice the 2-year yield minus the sum of the 3-month and 10-year yields.8 0.4 0. Slope and Curvature (Data-Based) Level Plot 12 10 8 Autocorrelation 6 4 2 0 -2 86 88 90 92 94 96 98 00 1.2 0.4 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Level Autocorrelation Level (Percent) Time Slope Plot 12 10 Slope (Percent) 8 Autocorrelation 6 4 2 0 -2 86 88 90 92 94 96 98 00 1.

5 6 25% 5.Figure 4 Median Data-Based Yield Curve with Pointwise Interquartile Range 9 8. Yield (Percent) .5 75% 8 7. we plot the median yield along with the twenty-fifth and seventyfifth percentiles.5 M edian 7 6.5 5 0 20 40 60 M aturity (M onths) 80 100 120 Notes to Figure 4: For each maturity.

Fixed 8t 0 .1 0 .1 6 λ E s t im a t e d t λ = 0 .0 2 0 Jan85 J u l8 7 Jan90 J u l9 2 T im e Jan95 J u l9 7 Jan00 J u l0 2 0 .0 3 5 λ E s t im a t e d t λ = 0 .0 8 5 0 .1 4 R M S E (P e rc e n t ) 0 .0 4 0 .Figure 5 RMSE of Yield Curve Residuals with Estimated vs.0 6 5 0 . we plot RMSE over time (averaged over maturity). .0 6 0 .0 4 0 .0 6 0 .0 8 0 . In both panels. and in the bottom panel we plot it by maturity (averaged over time).0 4 5 0 . the solid line corresponds to 8t estimated and the dotted line correspond to 8t fixed at 0.0 8 0 .0 6 0 9 t 0 .1 2 0 .0 7 5 R M S E (P e rc e n t ) 0 .0 6 0 9 t 0 20 40 60 M a t u rit y 80 100 120 Notes to Figure 5: In the top panel.0 7 0 .0 5 5 0 .0 5 0 .0609.

8 5.Figure 6 Actual (Data-Based) and Fitted (Model-Based) Average Yield Curve 7.6 6.2 7 6.2 6 Fitted Nelson-Siegel Actual 5.6 0 20 40 60 M aturity (M onths) 80 100 120 Notes to Figure 6: We show the actual average yield curve and the fitted average yield curve obtained by ˆ ˆ ˆ evaluating the Nelson-Siegel function at the mean values of $1t .4 7.8 Yield (Percent) 6. .4 6. $2t and $3t from Table 3.

8 5.8 9.1 9 8.6 5.2 5 4.6 6.7 9.2 Yield (Percent) Yield (Percent) 4. .9 7. together with actual yields.85 0 20 40 60 Maturity (Months) 80 100 120 0 20 40 60 M aturity (M onths) 80 100 120 Notes to Figure 7: We plot fitted yield curves for selected dates.95 6 5.8 6.3 9.5 Yield (Percent) Yield (Percent) 9.3 0 20 40 60 Maturity (Months) 80 100 120 Yield Curve on 5/30/1997 6. See text for details.5 9.4 5.Figure 7 Selected Fitted (Model-Based) Yield Curves Yield Curve on 3/31/1989 9.4 9.6 7.2 7.4 6.6 0 20 40 60 Maturity (Months) 80 100 120 7.4 7.8 9.05 Yield Curve on 8/31/1998 4.8 5 5.9 4.7 8 Yield Curve on 7/31/1989 7.9 7.

2000. 1985.01 . . See text for details.12 Notes to Figure 8: We plot residuals from Nelson-Siegel yield curves fitted month-by-month.Figure 8 Yield Curve Residuals.

Slope and Curvature (i.Figure 9 Model-Based Level. Data-Based Level. the slope as the difference between the 10year and 3-month yields..3 $3t Dotted Line: Curvature Notes to Figure 9: We define the level as the 10-year yield. and the curvature as the twice the 2-year yield minus the sum of the 3-month and 10-year yields. Estimated Factors) vs.e. . Slope and Curvature 14 12 10 8 6 4 86 88 90 92 94 96 98 00 ˆ Solid Line: $1t Dotted Line: Level 6 4 2 0 -2 86 88 90 92 94 96 98 00 ˆ Solid Line: & $2t Dotted Line: Slope 2 1 0 -1 -2 86 88 90 92 94 96 98 00 ˆ Solid Line: 0.

8 0.0 0.4 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Partial Autocorrelation ˆ Partial Autocorrelation of $1t 1.2 -0.0 0.4 0.2 -0.6 Autocorrelation 0.0 -0.4 0.0 0.0 0.0 0. $2t .2 0.8 0.4 0.2 0.8 0.0 -0.0 -0.2 0.6 Autocorrelation 0.0 -0.8 0.Figure 10 Sample Autocorrelations and Partial Autocorrelations of Estimated Factors ˆ Autocorrelation of $1t 1.6 Autocorrelation 0.6 0. .2 -0.4 0.2 0.2 0. along with Barlett’s approximate 95% confidence bands.4 0.0 -0.2 -0.2 -0.6 0.6 0.0 0.0 -0.8 0.8 Partial Autocorrelation 5 10 15 20 25 30 35 40 45 50 55 60 Displacement 0.4 5 10 15 20 25 30 35 40 45 50 55 60 Displacement ˆ Autocorrelation of $3t 1.2 0.4 5 10 15 20 25 30 35 40 45 50 55 60 Displacement ˆ Autocorrelation of $2t 1.4 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Partial Autocorrelation ˆ Partial Autocorrelation of $3t 1.2 -0.4 0.4 ˆ Partial Autocorrelation of $2t 1.4 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Notes to Figure 10: We plot the sample autocorrelations and partial autocorrelations of the three estimated ˆ ˆ ˆ factors $1t . and $3t .

.2 Autocorrelation of .1t .0 -0.3t } from univariate ˆ ˆ ˆ ˆ ˆ ˆ AR(1) models fit to the estimated factors { $1t .1 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Time Notes to Figure 11: In the left panel. we show time series plots of residuals { . and in the right panel we plot their sample autocorrelations.Figure 11 AR(1) Factor Model Residuals and Residual Autocorrelations Plot of .3t ˆ 3 2 Residual (Percent) 0. $2t .2 Autocorrelation of .3t ˆ 0.1 Time Plot of .0 -0.2t ˆ 3 2 Residual (Percent) 0. to a displacement of 60 months. $3t } .1t ˆ 3 2 Residual (Percent) 0.1 Autocorrelation 1 0 -1 -2 -3 86 88 90 92 94 96 98 00 -0.1 Autocorrelation 1 0 -1 -2 -3 86 88 90 92 94 96 98 00 -0. .0 -0.2t .1 5 10 15 20 25 30 35 40 45 50 55 60 Displacement Time Plot of .2 0.2 0. .2 Autocorrelation of . along with Bartlett’s approximate 95% confidence bands.1t ˆ 0.1 Autocorrelation 1 0 -1 -2 -3 86 88 90 92 94 96 98 00 -0.2 5 10 15 20 25 30 35 40 45 50 55 60 Displacement 0.2t ˆ 0.

Sign up to vote on this title
UsefulNot useful