Journal of Financial Economics IO (1982) 323-329.

North-Holland Publishing Company

DO FORECAST ERRORS OR TERM PREMIA REALLY MAKE THE

DIFFERENCE BETWEEN LONG AND SHORT RATES?

Richard STARTZ*

Received November 1981, final version received May 1982

Forward rates In the term structure of interest contain predictions of future spot rates plus
(possibly) term premia. Realized spot rates contain predicted spot rates plus forecast errors.
Under rational expectations forecast errors are not predictable. By forecasting spot rates using
publicly available information, bounds on the variation of forecast errors, and term premla are
obtalned. f-or one-month treasury bill rates, one to two thirds of the variation in the difference
between forward rates and realized spot rates is due to variation In term premia.

1. Introduction

It is fairly well established that long-term interest rates contain time-

varying term premia as well as forecasts of future short-term rates. Is the
variation in term premium Important? The natural way to pose such a
question is from the vantage of an investor, speculator, or capital budgetter
trying to extract a prediction of future short rates from the term structure.
Suppose one estimates the future short-term rate by the difference between
current long and short rates, perhaps adjusting further by a constant term
premium. The difference between this estimate and the short rate actually
realized in the future contains two sources of error: the markets error in
forecasting the future short rate and the error due to variation in the
premium embedded by the market in the long rate. This paper shows that
the variation in the premium is substantial compared with variation in the
markets forecast error.
The assumption that expectations are rational sets restrictions on the
joint stochastic processes generating spot rates, expected spot rates, forecast
errors, and term premia. In place of specific models of expectation and term
premium formation, these restrictions are used here to obtain a lower bound

*The author thanks G. William Schwert and an anonymous referee for both substantive and
expositional suggestions.

0304-405x/82/oooo-oooo /$02.75 0 1982 North-Holland

324 R. Srartz, Forecast errors and term premia in forward rates

on the time series variance of term premia and an upper bound on the
variance of forecast errors. The rational expectations restrictions are
insufficient to identify an actual time series for the term premium; however,
the restrictions are sufficient to identify a bound for the variance of the time
series of the term premium.

2. Population statistics

We can interpret the term structure as a set of spot rates and implied
forward rates. The spot rate realized next period, s, is the markets
expectation formed today, se, plus a forecast error, E. The implied forward
rate, f, is the expected spot rate plus a premium that the market chooses
today, P. Define the markets information set today, a set that obviously
includes se and P, as @, and assume that the market forms S rationally, in
the usual sense of se being a mathematical expectation, Restating all this in
algebraic form,

S-s=+&, (1)

f -sC+P, (2)

SC= E(s 1q. (3)

Rational expectations implies that E is uncorrelated with elements of the

information set. The following second moments are immediate:

var (s) = var (se) + var (E), (4)

var (s - f) = var (P) + var (s). (5)

Since spot and forward rates are observable, the statistics on the left of (4)
and (5) can be estimated. Shiller (1979) observed that a test for the pure

Suppose we want to examine the forward rate on a r-period bond beginning n--r periods in
the future, based on the current n-period long rate. Let ,R, be the yield on an n-period pure
discount bond issued in period t. Under certainty we have

(1 +,R,)=(l+ ,+,v,R,)(l +,R,-,Y.

We deline the realized spot rate and the forward rate as

s=log(l+ ,+,_,R,) and f=log(l+,RJ-log(l +,R, ,)-.

 R. Startz, Forecast errors and term premia in forward rates 325

expectations hypothesis, that P is zero or at least constant through time, is

to check whether var(s)> var(f), as it must be in the absence of a time-
varying risk premium.* Fama (1976) looked at the question of market
efficiency by making use of the fact that @ might be useful in predicting se
and P, but not E. Both Shiller and Fama find ample evidence that a premium
is present in forward rates. In his table 4, Fama present estimates of a simple
model of premium determination. The coefficients of determination there can
be compared with column (2) in the lower half of table 2 below. The evidence
below indicates that variation in the premium is even larger than Famas
specific model suggests.
Three underlying population statistics, var (se), var (P), and var (E), generate
the two observable statistics in (4) and (5). If one can only discover any one
of the three, the other two are immediately identified. Rational expectations
cannot identify any of the basic time-series per se. However, a simple
statistical filtering device identifies useful bounds on the Second moment
statistics.
If s is regressed on any subset of @, all the explained variation is due to se,
and none to E, since E is uncorrelated with @. The standard error of the
regression is an upper bound on the standard deviation of E. The following
estimators of the underlying population statistics follow immediately:

vir (E)= standard error of the regression squared [upper boundj (6)

vdr (se) = var (s) - vlr (E) [lower bounJl (7)

vir (P) = var (s -f) - vir (E) [lower bound] (8)

3. Sample statistics

Unconditional variances do not exist if the underlying series are non-

stationary. The sample series for s and f appear to be non-stationary. The
series s -1 and the estimated t: (the residuals from the filtering regression)
appear to be stationary, as shown by the autocorrelations reported in table 1.
Based on this evidence, variance bounds are reported below for the term
premia and forecast errors, but not for the expected spot rate.
With infinite data, one could calculate the bounding estimators (6) and (8)
with any desired (probabilistic) accuracy. Being restricted to historical time
series of finite length, one can only estimate (6) and (8). Some statement
about the precision of our point estimates is also required. In particular, the
time series correlation of the sample series must be taken into account. I
calculate approximate confidence bands using the methods suggested by

Singleton (1980) discusses hypothesis testing for these implied variance bounds
326 R. Star& Forecast errors and term premia in forward rates

Table 1
Sample autocorrelations from one-month bills.

Realized
spot rate Estimated
minus forecast
forward rate error
Lag (months) (s-f) (6)

1 0.36 - 0.03
2 0.22 -0.05
3 0.23 0.09
6 0.08 -0.11
9 0.21 0.06
12 0.25 0.13
15 - 0.02 -0.04
18 - 0.02 -0.02
21 0.04 - 0.03
24 0.10 0.15
27 - 0.03 0.01
30 -0.11 -0.07
33 -0.04 -0.01
36 0.03 0.15

Jenkins and Watts (1969, see in particular sections 5.3.1, 8.2.1, and app.
A9.1.) The covariance between variable Xi and variable Xj lagged u periods,
cij, is estimated by the usual procedure,

Cij(U)=~~~(Xi,-Xi)(X,.,+,-~j).
I 1

The covariance between any pair of these second-moment estimators is

given by

(cik(r) Cjdr) + cil(r) cjk(r)}, (10)

Actual application of formula (10) requires the acceptance of several

approximations. The formula is an asymptotic approximation, neglecting
terms of order l/T2 and smaller. In evaluating (lo), one necessarily must
substitute sample second moments for population second moments. Finally,
with a finite sample, the range of r must be restricted enough to allow
formula (9) to be applied with some accuracy. I restrict r to lie within plus
and minus 36 lags.
With infinite data the regression of s on a subset of @ should include every
variable in the publics information set that might reasonably contribute to
predicting the future spot rate. With time series of finite length enough right-
 R. Startz, Forecast errors and term premia injorward rates 327

hand side variables will produce an arbitrarily good lit. To avoid too good
a tit the information set included only lagged spot and forward rates and
only as many lags as required to yield residuals appearing to be white noise.
The same regression was then estimated for all maturities. Specifically, the
right-hand side variables are a constant, s_ r, s_ 2, L and f_ 1.3
Looking at the forward rates contained in, say, a 12-month treasury bill,
one might examine the projection for a one-month bill starting in 11 months,
an 11-month bill starting next month, or combinations in between. Table 2
presents results for long projections of one-month bills and one-month
projections of long bills.4 The tirst column gives the maturity of the bill, e.g.,
the first row reports results for the one-month-forward projection on one-
month rates. Column 2 reports variance of the premium as a percentage of
the variance of the forward deviation s-J Columns 3 and 4 report bounds on
standard deviations reported at simple annual interest rates (e.g., in row one
the typical deviation of the premium is 28 basis points) for the term premium
and the forecast error, respectively. Column 5 gives the sample mean of the
forecast spot rate for comparison.

4. Conclusions

Is the premium important? The second column answers the question. If

one uses the forward rate to forecast one-month rates, one-third to two-
thirds of the error is due to variations in the term premium, as a lower
bound. For forecasting next months long rate the result is less striking,
though as a lower bound ten percent may be worth some thought.
The evidence tells us something of the potential value of economic analysis
in the face of efficient financial markets. A planner interested in future short
rates would be well advised not to take todays implied forward rate as an
estimator. Even if it were granted that economic analysis cannot reduce the
forecast error, E, the determinants of the premium, P, are potentially
predictable. As the table shows, the variation in the premium is of the same
order of magnitude as the variation in the forecast error. Concomitant with
the large variation in the premium, one sees that typical market forecasting

3Note the date conventions. s_ , and f are contemporaneous quotattons at time t. The spot
rates are for the maturity of the bill being forecast, i.e.. one-month rates in the lirst part of table
2, one-, two-, live-, etc. month rates in the second part of table 2. R*s range from 0.96 to 0.30
for projections of one-month spot rates and from 0.96 to 0.92 for one-month-ahead projections.
4The data are for U.S. treasury bills, which are pure discount notes, and were developed by
Bildersee (1975). The sample period IS the same as that used by Fama (1976). monthly
observations from I/53 through 7/71 for the shortest bills. Longer bills have been generally
available only more recently, the shorter periods reported are 2/59 through 7/7l and IO/63
through 7/71. (I reran the 2-, 3-, and 6-month bill results over this last interval to ensure that
reported differences are not due to use of different periods. The results are approximately the
same as those reported. The percentage of forward deviatton due to the premium is marglnally
higher in the shorter periods.)
328 R. Startz, Forecast errors and term premia in forward rates

Table 2
Estimated variance bounds.

Standard deviations at
Percent of annual rates
forward -_ Sample
deviation due Term Forecast mean of Number
to premium premium error forecast Of

(lower bound) (lower bound) (upper bound) spot rate observations

(r + I)-month-ahead
forecast of
one-month rate
2 44.3 0.28 0.32 3.17 223
(7.0) (0.04) (0.02)
3 36.3 0.33 0.44 3.17 223
(8.2) (0.06) (0.03)
6 55.3 0.75 0.67 3.89 150
(8.5) (0.12) (0.07)
9 56.2 0.93 0.82 4.65 94
(11.3) (0.15) (0.15)
12 69.4 1.34 0.89 4.65 94
(10.4) (0.18) (0.19)

l-month-ahead
forecast of
(T + I)-month rate
2 44.3 0.28 0.32 3.17 223
(7.0) (0.04) (0.02)
3 21.8 0.17 0.31 3.34 223
(6.2) (0.03) (0.02)
6 12.4 0.1 I 0.29 4.28 150
(6.4) (0.03) (0.02)
9 15.6 0.14 0.33 5.16 94
(5.0) (0.03) (0.03)
12 15.1 0.14 0.34 5.20 94
(6.0) (0.03) (0.03)

Column 2 reports 100. var(P)/(var(P) + var (E)). Columns 335 report annual percentage rates.
Approxtmate standard errors, appearing in parentheses, are derived using eq. (10) to tind
vartances of the sample statistics and then applymg a tirst-order Taylor approxtmation, e.g..
var (a,) 2 var ((r:)/(4. ~7:).
The upper row marked 12 reports results for the I l-month-ahead forecast of l-month rates
contained in 12-month treasury bills. The lower row marked 12 reports results for the I-month-
ahead forecast of 1 l-month rates contained in 12-month treasury bills.

errors are much smaller than had been previously thought. For predicting
long rates into the relatively near future, simple use of the term structure is
fairly reasonable. For longer-term predictions, attention to changes in the
market premium is a must.
 R. Startz, Forecast errors and term premia in forward rates 329

References
Bildersee, John S., 1975, Some new bond indexes, Journal of Business 48, 506525.
Fama, Eugene F., 1976, Forward rates as predictors of future spot rates, Journal of Financial
Economics 3, 361-367.
Jenkins, Gwilym M. and Donald G. Watts, 1969, Spectral analysis and its applications (Holden-
Day, San Francisco, CA).
Shiller, Robert J., 1979, The volatility of long-term interest rates and expectations models of the
term structure, Journal of Political Economy 87, 1190-1219.
Singleton, Kenneth J., 1980, Expectations models of the term structure and imphed variance
bounds, Journal of Political Economy 88, 1159-I 176.