You are on page 1of 28

Financial Signal Processing: A Self Calibrating Model

Federal Reserve Bank of Chicago

By: Robert J. Elliott, William C. Hunter and Barbara M. Jamieson

WP 2000-21

Financial Signal Processing: A Self Calibrating Model

Robert J. Elliott Department of Mathematical Sciences, University of Alberta, Edmonton, Alberta, Canada T6G 2G1 William C. Hunter Federal Reserve Bank of Chicago, 230 South LaSalle St. Chicago, Illinois 60604 U.S.A. and Barbara M. Jamieson Department of Finance and Management Science University of Alberta Edmonton, Alberta, Canada T6G 2R6

September 2000

The views expressed here are those of the authors and not necessarily those of the Federal Reserve Bank of Chicago or the Federal Reserve System. Robert Elliott wishes to thank SSHRC for its support, and also the Federal Reserve Bank for its hospitality.

.A. Hunter(2) and Barbara M. Jamieson(3) (1) Department of Mathematical Sciences. Chicago. Elliott(1) William C. Edmonton. U. and also the Federal Reserve Bank for its hospitality. bond pricing. Department of Finance and Management Science University of Alberta. Robert Elliott wishes to thank SSHRC for its support. Canada T6G 2R6 (2) (3) Abstract.S.FINANCIAL SIGNAL PROCESSING: A SELF CALIBRATING MODEL Robert J. Alberta. Canada T6G 2G1 Federal Reserve Bank. Previous work on multifactor term structure models has proposed that the short rate process is a function of some unobserved diffusion process. so the model is selfcalibrating. Discretizing our model allows the use of signal processing techniques from Hidden Markov Models. Edmonton. Keywords: Term structure. This enables us to obtain explicit expressions for the prices of zero-coupon bonds and other securities. The estimation procedure is tested on a selection of U. Acknowledgments. Hidden Markov Model. Treasury bills and bonds.S. filtering. We consider a model in which the short rate process is a function of a Markov chain which represents the ‘state of the world’. 230 South LaSalle St. Alberta. University of Alberta. This means we can estimate not only the unobserved Markov chain but also the parameters of the model. Illinois 60604.

we shall follow their example and work under measure Q. Duffie and Kan (1993a. However. In fact Duffie and Kan suppose that Xt is given by a stochastic differential equation dXt = µ(Xt )dt + σ(Xt )dwt X0 ∈ D ⊂ Rn . P ) and a martingale measure is a probability measure Q equivalent to P such that the price processes of any security is a Q t r ds) 0 s t 0 martingale after normalization at each time t by exp ( = exp r(Xs )ds .1. This is not unreasonable as any diffusion can be approximated by a Markov chain. A comprehensive survey can be found in the paper of Duffie and Kan (1993a). (i. t ≥ 0. There. both single factor and multi-factor term structure models are described and a new model for the ‘short rate’ process rt is proposed under which rt is a function r(Xt ) of a ‘state process’ Xt which takes values in a subset D ⊂ Rn . Introduction The modelling of interest rates and term structure is a central problem in financial theory. Here we are assuming our processes are defined on a complete probability space (Ω.b) do not discuss the existence of an equivalent martingale measure.e. and others. µ(X) = A + BX). It is known from the work of Harrison and Kreps (1979). where µ and σσ are affine functions of X. Duffie and Kan then show that the yield process also has the form yt. F . that only technical conditions are required for the equivalence between the absence of arbitrage and the existence of an equivalent martingale measure. but assume such a measure Q exists. instead of assuming the short term rate r is a function of another diffusion process we shall suppose r is a function of a continuous time Markov chain. Discrete time 2 .τ = α(τ ) + β(τ ) · Xt .

. However. 0. . so with denoting transpose ei = (0. t t t pi = Q(Xt = ei ) = E[ Xt . eN }. . . Ingersoll. We do not need to specify a priori the dynamics of the short rate process. . or any space whatsoever. s2 . Suppose {Xt }. . is a finite state Markov chain on (Ω. The conclusion of the Pearson and Sun (1994) paper is that their data rejects the Cox. . . Maximum likelihood estimation of the Cox. ei ]. Short Term Rate Processes will be defined on a probability space (Ω. Q) with state space S = {s1 . Filtering methods provide a continual. . will denote expectation under where The distribution of Xt E[Xt ] = pt = (p1 . so that our model is ‘self calibrating’. Aggoun and Moore (1995) which provide not only recursive estimates of the Markov chain but also formulae for re-estimating all parameters of the model. will simplify the algebra. recursive up-date of optimal estimates in contrast to the static modelfitting of maximum likelihood. dt 3 . sN }. e2 . 0) ∈ RN . Ross model. . . . 1. 2. other than to say it is a Markov chain. the novel feature of this work is the application of new results from Elliott. p2 . without loss of generality we may identify the points in S with the unit vectors {e1 . . pN ).Markov chain models for term structure have been discussed by Pye (1966) and Zipkin (1993). Ross term structure model is carried out in the paper by Pearson and Sun (1994). . however. t ≥ 0. Here the points si may be points in RN . 0. for pricing purposes. We suppose our vectors are column vectors in RN . Q is an equivalent martingale measure. . F . . F . is then E This representation of the state space of X measure Q. . Ingersoll. Q) where. t We suppose this distribution evolves according to the Kolmogorov equation dpt = Apt . Consequently. . . . our application of Hidden Markov filtering and estimation techniques appears new.

.1. Bond Dynamics Suppose we have a security paying u(XT ) at time T ≥ t. e1 . that is. Consequently. h2 . Taking u = 1 we obtain the price of a zero-coupon bond maturing at time T : T pt. . 1 ≤ i. We suppose the short rate process rt rt = r(Xt ) = r. . Hypothesis 1. Then.T = E[exp − The yield for such a bond is then yt. as X is a Markov process. the price of a E[exp − t r(Xs )ds u|Ft] = E[exp − t r(Xs )ds u|Xt ]. because we are working under the equivalent martingale measure security paying u at time T ≥ t is given by T T Q.T . j ≤ N. This would still allow estimation of is a function of Xt . to be time varying. filtration generated by X.T = − t r(Xs )ds |Xt ]. so h(Xt ) = h. though this of the Markov chain is one of the unit vectors. is just given by a vector (h1 . Note all expectations are under Q. if A = (aji ). any real valued function of Xt . complete. eN . 1 (T −t) log pt. we could take r X but would complicate the parameter estimation of the model. . Xt for some vector r ∈ RN . Write {Ft } for the right continuous. hN ) = h. . As noted above. . . where the brackets Again. . Xt denote the scalar product in RN .here A is a ‘Q-matrix’. Recall that at any time t the state Xt to be time varying. say h(Xt ). that is. N j=1 aji = 1 and aji ≥ 0 if i = j. its 4 . 3. e2 . so the dynamics and estimates derived below are also under measure Q. We could take the components aji would complicate their estimation.

Xt . Xs + φs . Further. Xs . r(Xs )ds u(XT )|Ft ]. φ2 . consequently dF (Xs . t) = F (X0 .1) + 0 t exp − exp − r(Xv )dv s dφs . Xs ds ds (3.2) 5 . Differentiating we have the following Itˆ representation: o t t s exp − 0 r(Xs )ds F (Xt . . the semimartingale form of the Markov chain is: t Xt = X0 + 0 AXv dv + Mt (3. AXs ds φs . t) = E[exp − t r(Xs )ds u(XT )|Xt ]. 0) + t 0 − r(Xs ) exp (− s 0 0 r(Xv )dv φs . Consequently there is a function φt = (φ1 . Now t t exp − 0 r(Xs )ds F (Xt . t t t where φi = F (ei . . dMs . . s) = φs . t) = exp − 0 r(Xs )ds T 0 φt . Xt = E[exp − and so is a (Q. Xs + φs . dXs . as above. s) = dφs . + 0 0 r(Xv )dv We are using here the representation F (Xs . such that t F (Xt . φN ) ∈ RN . t) = φt . . Ft ) martingale. This is a function of Xt .price at time t is T F (Xt . t).

where diag r is the matrix with r on the diagonal.where {Mt } is a (Q. AXs ] = 0. . .2) see Elliott.1) is a martingale. . Xs + φs . . That is s exp − 0 r(Xu )du [−r(Xs ) φs . . φt . . therefore. Xt = e−B(T −t) u.3). Ft ) martingale. . . the bounded variation terms on the right side of (3. 6 . . so φt = e−B(T −t) u and the price at time t ≤ T security paying u(XT ) at time T is of a for all Xs . rN ) dφs . Xs − diag r · φs .1) must be the identically zero process. with A∗ denoting the transpose of matrix A. 1. ds and r(Xs ) φs . from (3. Therefore. dφs . XT . where u(XT ) = u. Aggoun and Moore (1995). Xt . m. Xs + where r = (r1 . i = 1. 1) so its price at 4. r2 . .3) Now r(Xs ) = r. . Xs . . . Filtering and Model Estimation We have shown the zero-coupon bond price which expires at time t+τi . uN ) . φs is given by the system of equations dφt = (diag r − A∗ )φt dt with terminal condition φT = u = (u1 . . A zero-coupon bond corresponds to taking time t ≤ T is F (Xt . t) = e−B(T −t)1 . Xs = 0 ds Consequently. For the derivation of (3. The left side of (3. Xt . Xs + A∗ φs . u = 1 = (1. . . Write B = diag r − A∗ . Xs diag r · φs . u2 . Xs = (3.

. . that is. Xt 7 . . t) = − log e−Bτi 1 . . (Note that theoretically this model would allow bond prices greater than one. . . such as those with ‘long tails’. . . . 2. . . . t) and the corresponding prices F i (Xt . t2 . . . .Xt bi t . . . where bi ∼ N (0. k. σ i (2). . − 1 1 log F i (Xt . τi τi t = t1 . F i (Xt . i = 1. . . .) In the estimation theory developed below the Gaussian random variables b. t) > 0 we suppose the t1 . i = 1. Xt τi τi = g i . . . noise is multiplicative. . . could be replaced by other random variables. . Xt . m. .is F (Xt . Our hypothesis now is that the give rise to bond prices state process Xt X process represents unobserved factors which would in some ‘ideal’ world. t) − bt . . t t = tj . t) Rather we observe these prices τ 1 . σi (N ) . t) = E[exp − i t+τi t r(Xs )ds|Xt ] = e−Bτi 1 . Write σ i (j) = −σ i (j)/τi and σ i = σ i (1). . t) F i (Xt . we suppose the are not observed directly. σ i (1). tk . we suppose we observe the yield values in additive noise: i yt = − 1 σ i . tk . t2 . . . τm . . σ i (2). σ i (N ) . Xt i log F i (Xt . . . . That is. . t)e and σ i = σ i . 1). Also. . . . at discrete times in noise for the different times to maturity As F i (Xt . . . m. in practice these are not observed. we suppose that what we actually observe are the quantities F i (Xt . representing noise. . j = 1. . Equivalently.

1. σ i are given The filtering and estimation algorithms for the parameters P = (pij ). . Q) martingale increment. that is. X bi . Suppose the observation times t1 ≤ t2 ≤ t3 ≤ . . .g. . . 8 . the choice of the value of parameter N is discussed.S. . . and the results of parameter estimation and yield prediction are presented and evaluated. =t . +M is an (F . m.S. Treasury bills and 10-year and 30-year U. . . parameter N. e. 5. i = 1. in the Appendix. 2.for g i ∈ RN . Here g i = g i (1). 5. ‘good’ and ‘bad’. . tj+1 − tj = s > 0. 2. bonds. In what follows. Then we have a discrete time version of the state process X = Xt . . which represents the size of the state space of the Markov chain. . X + σi . . a value is assigned to N which can be thought to represent the number of states of the world. the data are described. . . k. is the only parameter which is not estimated. k. Application The results of the paper were applied in an example using data on the yields of 3-month and 6-month U. g i (2). g i (N ) − 1 τi and g i (j) = are equally log e−Bτi 1 . . . . . g i . 1 ≤ i ≤ m. . and write eAs = P. = 1. . The multivariate observation process y has dynamics yi = g i. . The Choice of N In the estimation procedure proposed above. with X = PX where M −1 = 1. ej . spaced. Rather. . . .

a state space of size two is typically assumed. as measured by standard deviations. is an important problem which has been considered in the literature. Thus. there is evidence. Table 1 provides descriptive statistics for the yield data. a number of proposals have been advanced to address it. The sample period ran from January 17. see Hamilton (1994). 1994) for example. if we consider the two subperiods. a smaller value of sensitive to the value assigned to N N increases in size. However. other than to compare the results obtained when N is assigned different values. Although this problem cannot be resolved using the likelihood ratio test. this provides further support for assigning N 5. unless our results are and superior for large N. therefore. in this case). Also.2. From a numerical is. before mid-1994 and after mid-1994.S. yield volatilities increase with maturity. it is interesting to note that in the regime-switching model. rather than larger. the earlier period coincides with a relatively low short-term rate regime and the later period 1 For a smaller. The model that we propose here requires the estimation of N 2 + 2mN parameters. that suggests that the model is overfitting the data. value. more on this issue.S. coupon bonds. for a particular data set. the dimensionality of the model increases rapidly as N point of view. 1992 to June 21. However. where m equals the number of securities being considered (four. Treasury bills and 10-year and 30-year U. 9 . preferable. Data Description We assembled a data set consisting of 270 weekly observations on the yields of 3month and 6-month U. that is. yield volatilities.1 In this paper. discussed in Hamilton (1988. If the entire period is considered. a reverse pattern is observed. demonstrating that the yield curve was typically upward sloping during the sample period. The data were compiled by the Royal Bank of Canada and published in The Financial Post.The determination of the optimal value of N. pages 698-699. we do not explore this issue further. 1997. in which the state or regime of a time series process is modelled as a Markov chain. appear to decline with maturity. Also. reported in what follows.

. was run on the data set. A. m. AXl ˆ where Xl = E[Xl | yl ]. . predicted yields for the following week were obtained for each of the four securities considered. . a value of two was assigned to parameter of observations consists of 100 yield vectors. price estimates were obtained. k. . . All entries in the transition matrix. . and for the matrices g and σ. . the first group consisting of 100 yield vectors and subsequent groups consisting of 10 yield vectors each. . At the end of each pass through the data. . respectively. and l = 1. written to implement the estimation procedure proposed in Section 4. 2.with a relatively high short-term rate regime. We then regressed actual yields on predicted yields for each 10 . Note that the columns of the g and σ matrices and the price vector correspond. 5.and 30-year bonds. Analysis and Results A computer program. we calculated predicted yields using the formula: i ˆ E[yl+1 | y1 .3. to the 3. that is. . Estimated prices are also reported. were assigned an initial value of 1/N. To assess the predictive performance of the model. for E[X0 ]. . parameter estimates were updated using the formulas given in the paper. To start. . At the end of each of the 18 passes through the data. a value of N. N Because the first group in excess of 15 would be inappropriate since it would result in a situation where the number of parameters to be estimated exceeds the number of observations. As well. i = 1. Table 2 also reports the re-estimated values of these parameters after the eighteenth pass through the data. The data were processed in 18 groups of observations. . yl ] = g i . Table 2 gives the initial values that were assumed for the distribution of the state of the Markov chain. .and 6-month Treasury bills and the 10. . .

we divided the data in half. and a regression coefficient.and 6month Treasury bills and the 10. the standard error of the estimate of α is relatively large in all four cases. the results suggest that the first criterion. To apply this method. Figures 1 to 4 provide plots of actual yields and predicted yields for the 3. close to zero.and 6-month Treasury bills indicate that the test for serially uncorrelated residuals is inconclusive. residuals were recorded. an intercept.96. These values were plotted.of the securities in turn. the predicted yields were then compared to the actual yields. For each of the securities in turn. They also give standard error bands for the predicted yields. derived using the method of cross-validation. Table 3 reports the results. using the model: Actual yield = α + β ∗ Predicted yield + . At the end of each of the 18 passes through the data. respectively. The second set of 135 observations was similarly processed and a set of 18 vectors of predicted yields obtained. close to one. conditional unbiasedness. and (3) a low residual standard error. For each of the four securities considered. while the results for the 3. Also. However. the first group consisting of 50 observations and subsequent groups consisting of 5 yield vectors each. β. (2) serially uncorrelated residuals. The first 135 observations were processed in 18 groups. along with the corresponding actual yields. evidence for the 10. α. The use of cross-validation to determine the standard error bands was deemed nec11 . Standard error bands around the predicted yields were derived by adding and subtracting to each predicted yield an amount equal to the product of a critical value. and the standard deviation of the residuals derived using the first 135 observations. a yield prediction for the following week was obtained. in Figures 1 to 4. is satisfied.and 30-year bonds supports the hypothesis that residuals are not autocorrelated. The regression results obtained were then assessed on the basis of the following criteria proposed by Fama and Gibbons (1984): (1) conditional unbiasedness. that is.and 30-year bonds. 1. and the standard deviation of the residuals was calculated.

Empirical work on bond prices show that a small state space for X is better and that our model predicts yields quite well. . AXl ]2 . Tables 4 and 5 report results. . . .essary because evidence suggested that the model was overfitting the data. i = 1. . 2. The estimated values reported here are similar in magnitude to those given earlier in Table 2. k. These results lead to conclusions identical to those reached earlier for Table 3. . Although possible overfitting of the data militates against increasing the value of N. Thus. AXl ± 1. . enable us to estimate not only X but also the parameters of the model. using new results presented in Elliott. Conclusion Our model of the short rate process gives rise to expressions for yields which incorporate two random components: a Markov chain X and a Gaussian noise term b. 6. . Tables 6 to 8 report results. . it was deemed appropriate to derive the standard error bands using cross-validation methods. This evidence included the observations that our data on Treasury bill and bond yields are not highly variable. Aggoun and Moore (1995). m and l = 1. the analysis was repeated for some alternative N values to determine whether the results appear sensitive to changes in the value of N. . calculated using the following formula: ˆ g i . of the type given in Table 3. respectively. as reported in the σ matrix. 12 .96 ˆ ˆ ˆ g i (diag AXl )g i + σ i (diag AXl )σ i − [ g i . Filtering techniques. . for the cases N = 4. our estimates of yield volatility. and standard error bands for predicted yields. of the type given in Table 2. are small in value. and N = 9. for the cases N = 4 and N = 6. . respectively. are too narrow in width. N = 6.

381-408. Rational expectations econometric analysis of changes in regime – An investigation of the term structure of interest rates. Duffie and R. Journal of Economic Dynamics and Control 12 (1988). of Econ. = 1. Sun (1994). 2. New Jersey. .References D.J. . Jour. 327-348. having dynamics X = PX −1 +M and a multivariate observation process Y Y i = g i . Series A. Math. J. P. Exploiting the conditional density in estimating the term structure: an application to the Cox. . 1994. L. Moore (1995). Duffie and R. 29. Kreps (1979).J. R. Trans. Applications of Mathematics. A Markov model of term structure. Quarterly Jour. Zipkin (1993). Appendix With a state process X . Mortgages and Markov chains: a simplified evaluation model.B. R. . = 1. therefore. M. with components 1 ≤ i ≤ M. . Pearson and T-S. X + σ i . Jour. 683-691. 1279-1304. Hidden Markov Models: Estimation and Control.D. . 379-406.D. New York. Gibbons (1984). Theory 20. of Economics 25. Kan (1994). Automatica 30. 1399-1408. J. Time Series Analysis. Pye (1966).D. Eugene F. we are. Springer-Verlag. London. Hamilton. Ingersoll and Ross model. 385-423. A yield factor model of interest rates. 2. Elliott (1994). 60-72. Martingales and arbitrage in multiperiod security markets. G. D. Harrison and D. N. 577-586. Phil. Vol. of Finance 49. Elliott. Princeton. Aggoun and J. Kan (1996). Fama and Michael R. A comparison of inflation forecasts. Exact adaptive filters for Markov chains observed in gaussian noise. Princeton University Press. Hamilton. in a situation analogous to the Hidden Markov Models discussed in 13 . 347. Journal of Monetary Economics 13. Management Science 39. . Finance 6. Royal Society. Multi-factor term structure models. X bi .

X φ(y i ) i m γ = i=1 γi and 14 . X γ = σ i . . random variables. . F ) such that under Q: a) X. 2. and b) the observations have the form mentioned in Section 8 of Elliott (1993). (see also the book by Elliott.i. The differences are that a) the observation process is multidimensional.Elliott (1993). as discussed in Section 7 of Elliott (1993). Following Elliott (1993).d. and b) yi . 1) i. 1) random variables. we recall that the analysis takes place under a probability measure Q for which the {y i } are i. .i. . where E[Mn |Fn−1 ] = 0.d. . so that Xn = P Xn−1 + Mn . is a sequence of N (0.) In fact we suppose we have a probability measure Q on (Ω. X )/ σ i . N (0. . 1 ≤ i ≤ m. 2. (here E denotes expectation under Q). Aggoun and Moore (1995)). is a Markov chain with transition matrix P. . = 1. = 1. Write φ(x) = (2π)−1/2 exp (−x2 /2) and set φ (y i − g i . (Note this change of measure is a mathematical artifact and is different to the equivalent martingale measure Q which gives rise to the prices.

A version of Bayes’ Theorem (see [4]) implies E[Λ H |Y ] E[Λ |Y ] σ (H ) = .i. However. (yi − gi . say. X bi . as in Elliott (1994). Write {G } for the complete filtration generated by the processes X will denote the complete filtration generated by y. σ (1) E[H |Y ] = (A. that under Q N (0.d. 15 given .X ) σi . If {H } is any {G } adapted sequence we write σn (Hn ) = E[Λn Hn |Yn ]. so the y i give noisy observations of the state X. 1) random variables. Q is an easier measure mathematically with which to work.X it can be shown. under Q yi = g i. {Y } If we define bi = the bi are i. The probability measure Q can be defined by putting dQ |Gn = Λn . X + σi . dQ and y.n Λn = =1 γ for n ≥ 1. that is.1) where σ (H ) = E[Λ H |Y ] is an unnormalized conditional expectation of H Y.

ej = 1 = j=1 N E[Λn−1 Γj (yn )ej P Xn−1 + Mn . ei Γi (yn )ei N + σn−1 (Xn−1 ). ej |Yn ] = j=1 P σn−1 (Xn−1 ). . er Γs (yn )psr es . er rs X .Write Γi (yn ) = m j=1 j φ (yn −gj (i) /σj (i) j σj (i)φ(yn ) . then Nn is the number of jumps from state er to es up to time n. . ΓN (yn )) on the diagonal.2) = Γ(yn )P σn−1 (Xn−1 ) where Γ(yn ) is the diagonal matrix with (Γ1 (yn ). Suprs pose Nn = n X =1 −1 . Similar calculations give N rs σn (Nn Xn ) = i=1 rs P σn−1 (Nn−1 Xn−1 ). r With Jn = n X =1 −1 . . ej Γj (yn )ej (A. In particular consider σn (Xn ) = E[Λn Xn |Yn ] N = E Λn−1 γ n Xn j=1 N Xn . With f(y) a function of y and n Gr (f) n = =1 X . er f(y ) 16 . ei Γi (yn )ei + σn−1 (Xn−1 ). er i=1 Γi (yn )pir ei . es . er . ej |Yn because j=1 N Xn . N r σn (Jn Xn ) = i=1 r P σn−1 (Jn−1 Xn−1 ). . the occupation time in er .

the normalizing factor in (4. n As noted in Elliott (1994). 1 = E[Λn |Yn ]. rs r Therefore. the above formulae give expressions for σn (Nn ). summing the components. er i=1 Γi (yn )pir f(yn )ei . n n Also.3) j j σ i (j) = σn (Jn )−1 σn Gj ((y i )2 ) − 2g i (j)σn Gj (y i ) + g i (j)2 σn (Jn ) . the above estimates enable us to re-estimate the and σ i (j) at time n as: g i (j) = g i . ei Γi (yn )ei n−1 N + σn−1 (Xn−1 ). closed form recursions are obtained. Following Elliott (1994). unlike Hn alone. 1 = σn (Hn Xn . the transition probabilities in the matrix P = (psr ) can be re-estimated at 17 . we consider Hn Xn because. ej =− 1 log e−Bτi .1) is σn (1) = σn (Xn ). 1 ). σn (Jn ) and σn (Gr (f)). ej τi 1 =− log E exp − τi σn Gj (y i ) n = j σn (Jn ) g i (j) τi 0 r(Xs )ds |X0 = ej (A. Also. Now for any scalar process Hn σn (Hn ) = σn (Hn Xn ).N σn Gr (f)Xn n = i=1 P σn−1 Gr (f)Xn−1 .

if t = t so Xt = X E[F i (X .3).time n by rs r psr = σn (Nn )/σn (Jn ). t )|Y ] = e−Bτi 1 . our algorithms give a recursive filter for the unobserved chain.2) and σ (1) = σ ( X . τi − g i (j)τi . (and almost any process can be approximated by a Markov chain). − so the j th 1 log e−Bτi 1 . E[X |Y ] and E[X |Y ] = σ (X ) . For example. σ (X ) is given recursively by (A. t) = E[exp − t+τi t r(Xs )ds |Xt ] = e−Bτi 1 . ej = g i (j). Consequently. σ (1) Here. The components g i (j) are component of e−Bτi 1 is exp re-estimated in our model by equation (A. with F i (Xt . Furthermore. This in turn allows the re-estimation of the model parameters. 1 ) is the sum of the components of σ (X ). 18 . which include the yield prices for varying maturities. if we accept the model is reasonable. Xt .

54 Note: ’STD’ denotes standard deviation.79 6.35 6-month 4.46 1.02 3.01 3.29 0.27 0.48 0.70 6.61 6.97 0.47 5.42 0.07 0.57 7.37 10-year 6.TABLE 1: Descriptive Statistics for Treasury Yields 1/92-6/97 Maturity Mean STD 1/92-6/94 Mean STD 7/94-6/97 Mean STD 3-month 4.02 0.30 1.61 0. 19 .71 0.12 0.61 30-year 7.50 5.

51 4.97.13 5.89 0.50 −0.50 σ matrix: 0.90.50 0.01 3.50 0.50 0.50 0.00 0.07 σ matrix: 0.91 0.51.50 0.05 0.94 5.97 2.TABLE 2: Starting values and parameter estimates – 3.50.58 Estmated prices: (0.89 1. 20 .05 0.00 −0.50 0.07 0.11 0.07 0.50 0.50 After the eighteenth pass: E[X18 | 18 ] vector: (0.50) A matrix: 0.50 0.10) A matrix: 0.07 0.09 g matrix: 0.50 −0. 0.05 0. 0.and 6month Treasury bills and 10. 0.24 0.50 g matrix: 0. 0.and 30-year bonds.50 −0. N = 2 E[X0 ] vector: (0. 0.99.50 0.05 0.50 0.00 0.00 0.13) Note: All entries in the σ matrix that was derived after the eighteenth pass are to be multiplied by 10−6 .

08) β 0.96) 1.73 (0.21 (0.94 1.TABLE 3: Regressions of actual yields on predicted yields – 3.11 (1.88 0.15) R-squared 0.97 (0.93 0.84 (0.32 Note: The numbers in parentheses are the standard errors of the corresponding parameter estimates.97 0.89 (0.and 30-year bonds.06) 0.95 (0. N = 2 Term to maturity of security Parameter 3-month 6-month 10-year 30-year α 0.34 (0. 21 . ’s’ denotes the residual standard error.33 0.14 0.21 1.and 6-month Treasury bills and 10.22) 0.14) 0.19 0.33) 0.71 0.17 1.65 Durbin-Watson D statistic s 1.05) 0.

84 4.04.88 0.07 0.03 0.89. N = 4 Starting values: E[X0 ] vector: (0.73 3.25 −0.84 1. 22 .51 4.and 6-month Treasury bills and 10.00 0.04 0.07 0.25 0.25 0.50 After the eighteenth pass: E[X18 | 18 ] vector: (0.05 0.25) 0.00 0.37 0.50 0.51. 0.61 3. 0.01 0.25 −0.25.25 −0.99.97.96 2.05 0.25.25 0.01 0. 0.25 0.05 0.07 0.50 0.85 4.00 0.00 0. 0.50 0. 0.and 30-year bonds.90 0.00 g matrix: 0.25 0.78 4.07 0.04 0.50 0.04 0.03 0.25 0.25 0.00 −0.25 −0.18 Estimated prices: (0.05 0.83 1. 0.04) 0.04.04 0.05 0.25 σ matrix: 0.04 0.07 0.04 0.50 0.07 0.50 0.25 0.13) Note: All entries in the σ matrix that was derived after the eighteenth pass are to be multiplied by 10−6 .13 5.28 σ matrix: 0. 0.TABLE 4: Starting values and parameter estimates – 3. 0.84 1.07 0.05 0.25.92 0.25 0.50 0.05 0.91 A matrix: 0.03 0.00 0.76 3.34 0.25 −0.05 0.00 −0. 0.25 −0.07 g matrix: 0.05 0.

33 Estimated prices: (0.11 0.84 4.05 0.07 0. 0.26 0.05 0.07 0.42 4.96 2.13) Note: All entries in the σ matrix that was derived after the eighteenth pass are to be multiplied by 10−6 .07 0.07 0.51.12 5. 0.52 0.07 0.09 0.07 0.and 30-year bonds.05 0.05 0.85 4.01 0.and 6-month Treasury bills and 10.07 0.95 2.48 4.05 0.05 0.07 0.05 0.05 0.49 4.10 5.TABLE 5: Parameter estimates – 3.05 0.99.96 2.76 3. N = 6 After the eighteenth pass: 0.85 1.96 2.97.07 0.07 g matrix: 0.13 5.05 0.49 4.77 3.84 1.12 5. 23 .07 0. 0.05 0.37 σ matrix: 0.07 0.

35 (0.96) 1.93 0.14) 0.75 (0.22 1.08) β 0. 24 .TABLE 6: Regressions of actual yields on predicted yields – 3and 6-month Treasury bills and 10.15) R-squared 0.97 0.18 1.35 0. ’s’ denotes the residual standard error.97 (0.71 0.95 1.94 (0.89 0.12 (1.14 0.89 (0.05) 0.22) 0.19 0.65 Durbin-Watson D statistic s 1.32) 0.21 (0.06) 0.32 Note: The numbers in parentheses are the standard errors of the corresponding parameter estimates.and 30-year bonds. N = 4 Term to maturity of security Parameter 3-month 6-month 10-year 30-year α 0.84 (0.

32) 0.22) 0. ’s’ denotes the residual standard error.75 (0.35 (0.14) 0.19 0.06) 0.19 1.TABLE 7: Regressions of actual yields on predicted yields – 3.96 1.35 0.05) 0.89 (0.97 (0.94 (0.32 Note: The numbers in parentheses are the standard errors of the corresponding parameter estimates.71 0.15) R-squared 0.14 0.97 0.22 (0. 25 . N = 6 Term to maturity of security Parameter 3-month 6-month 10-year 30-year α 0.and 6-month Treasury bills and 10.89 0.08) β 0.65 Durbin-Watson D statistic s 1.and 30-year bonds.12 (1.93 0.96) 1.84 (0.23 1.

89 0.84 (0.06) 0.32) 0.04) 0.and 6-month Treasury bills and 10.19 0.07) β 0.20 (0. 26 .97 0.14 0.32 (0.14) 0.89 (0.TABLE 8: Regressions of actual yields on predicted yields – 3.74 (0.71 0.97 (0.18 1. ’s’ denotes the residual standard error. N = 9 Term to maturity of security Parameter 3-month 6-month 10-year 30-year α 0.22) 0.95 1.12 (1.and 30-year bonds.65 Durbin-Watson D statistic s 1.21 1.34 0.15) R-squared 0.95 (0.95) 1.32 Note: The numbers in parentheses are the standard errors of the corresponding parameter estimates.93 0.