MODELLING AUSTRALIAN BANK BILL RATES : A KALMAN FILTER APPROACH By Ramaprasad Bhar School of Finance and Economics University of Technology, Sydney P.O. Box 222, Lindfield, NSW 2070 Tel. 02 330 5422 Fax. 02 330 5515 ABSTRACT This paper examines the applicability of the Kalman Filter technique to forecast future spot interest rates, based upon the expectation hypothesis of the term structure of interest rates, in the Australian bank bill market. In this approach, regression estimates are based on the last period's estimate together with data from the current period. In contrast to constant parameter models, this allows effective use of information underlying the process driving the evolution of the parameters. For the period tested, forecasting accuracy of such a time-varying parameter model shows marked improvement over a constant parameter model. Acknowledgements: I am particularly indebted to one anonymous referee for very insightful comments on the earlier versions of the paper, and would like to thank the editor, Rob Brown, for helpful suggestions. Australian Bank Bill Rate: Kalman Filter Approach 1. INTRODUCTION: A number of recent research studies have found that Treasury Bill futures provide a better indication of future spot rates than the forward rates implied by the current term structure of interest rates. Notable among these studies are Cole, Impson and Reichenstein [1991], Hafer, Hein and MacDonald [1992], Kamara [1990] and MacDonald and Hein [1989]. Studies like Fama [1984], Hamburger and Platt [1975] and Kane [1983] suggest that the poor performance of the implied forward rate as a predictor of the future spot rate may be due to several possibilities such as the existence of a risk premium, policy changes or, possibly, inadequate model specification. Use of futures market data to forecast future spot rates is restricted to the futures contract delivery periods. Another approach, in this respect, is to regress future spot rates (ex-post basis) on some period ahead forward rates implicit in the term structure of interest rates. This Ordinary Least Square (OLS) method assumes constancy of parameters over the sample period and this assumption is a potential source of model mis-specification, as shown by Chiang and Kahl [1991]. From the forecasting point of view, this may lead to the least cost solution but may result in suboptimal forecast error. In this study standard statistical techniques are applied to establish time variation of the regression parameters in the Australian bank bill market. To account for this time variation of the parameters in the model, the Kalman Filter technique is employed as an adaptive method of estimation and forecasting. The result shows marked improvement in forecasting ability over the time invariant parameter model. The structure of the paper is as follows. Section 2 describes the Australian bank bill market and in Section 3 alternative models are developed. Data used in the model estimation and forecasting 2 Australian Bank Bill Rate: Kalman Filter Approach are described in Section 4 and the analysis of result follows in Section 5. The paper is then concluded in Section 6. 2. THE AUSTRALIAN BANK BILL MARKET: The physical bank accepted bills market is a significant short-term financial market in Australia. These bills are issued on a discount basis for a period of up to 180 days but the 90-day market is the most active. These bills have a face value of $500,000 and are usually traded in lots of $5 million. Rates are quoted on yield p.a. basis with 365 days in a year. The relationship between the face value (F) and the market value (P) is given by, P = (365 * F)/(365 + yield * days to maturity) (1) The data in Table 1 show the average of weekly assets ( bills and certificate of deposits ) of the authorised dealers between the years 1981 and 1992. The relative popularity of bills over CDs is evident until the late 1980s. The introduction of interest rate futures contracts based upon bank bills also reflects the importance of this instrument in the Australian financial market. Carew [1991, page 117], however, points out that the changes in the statutory reserve deposit (SRD) requirement introduced in 1988, may shift the relative importance towards CDs. There is some indication of this in Table 1 during the years 1990 - 1992. Carew [1991, page 118] also argues that "... the substantial volume of bills lines written for several years and still outstanding means that bills of exchange will be around for some time.". 3. MODEL BASED ON THE EXPECTATION HYPOTHESIS: The expectation hypothesis argues that a forward interest rate corresponding to a certain period is equal to the expected future spot interest rate for that period. Muth [1961] identifies several 3 Australian Bank Bill Rate: Kalman Filter Approach key properties for this expectation to be "rational expectation". One of them is the property of unbiasedness. To model this expectation, following Friedman [1980], the bank bill rate is regressed on k-period ahead forecast: r t = α + β t-k f t + ε 1,t (2) where r t is the actual bill rate at time t and t-k f t is the forecast of r t made at t-k. α and β are parameters of the model and ε 1 is the usual regression error term. Equation (2) is referred to as the time invariant expectation model in this paper. The unbiasedness property is examined by the null hypothesis: H0: (α,β) = (0,1). Also, the residuals in the regression equation should be free from serial correlation. This joint test of the parameters is equivalent to testing the pure expectations hypothesis. An implicit assumption in this model is that the behaviour of the parameters is time invariant. So, failure to reject the null hypothesis may imply that the estimated parameters are sensitive to the sample period used. For the model to be of practical use, it should be capable of capturing the dynamics of market behaviour over time. One way to achieve this would be to employ an adaptive filtering technique where the estimates of the parameters at time t would be related to the estimate of the previous period in some simple way. This is an heuristic approach and lacks a sound theoretical foundation since it does not take into account the underlying stochastic process describing the change (see for more detail Kahl and Ledolter [1983]). In order to apply a more rigorous technique the parameters are explicitly allowed to follow a random process over time and the resulting method is referred to as a Kalman Filter. A Kalman Filter commonly refers to estimation of state space models where there are two parts, i) the transition equation and ii) the measurement equation. The transition equation describes the evolution of the state variables (i.e. the parameters) and the measurement equation describes how 4 Australian Bank Bill Rate: Kalman Filter Approach the observations are actually generated from the state variables. Regression estimates for each time period in this case are based upon the previous period's estimates and data up to and including the current time period. The model is defined with the following equations: r t = α t + β t t-k f t + ξ t where ξ t ∼ N (0, σ2 ) (3) α t = α t-1 + η1 t and β t = β t-1 + η2 t , where (4) η1 t ∼ N ( 0, σ12 ) and η2 t ∼ N ( 0, σ22 ) , (5) and the initial values of α and β are assumed to follow, α0 ~ N( 0,σα2 ) and β0 ~ N( 0,σβ2 ). The variables r t and t-k f t are as defined in relation to equation (2) and ξ and η are classical well behaved disturbances associated with the measurement equation and transition equations respectively. Prior values of the coefficients and the variances are usually obtained from a regression of the initial observations of the sample. Appropriate starting values or initial conditions are crucial in Kalman Filter implementation. Harvey [1990, pp. 121-122] points out that when the transition equation is non-stationary (as it is in this study) the initial distribution of the state variables should be specified in terms of diffuse prior. Harvey [1990, pp. 121-122] also suggests that by setting σα2 and σβ2 to a large but finite number, a good approximation can be obtained. In the implementation of the Kalman Filter in this paper, both σα2 and σβ2 are set to 10000. The time invariant model corresponds to the special case where σ12 = σ22 = 0. The Kalman Filter1 approach should, thus, provide a much improved estimate of the relationship between the future spot rate and the forward rate. In modelling interest rates, a random walk process is also suggested by various researchers. For example, Berger and Craine [1989] show that "the random walk model might be a good 1The method of estimation in Kalman Filter is maximum likelihood conditional on the data observed up to that point. In that sense it can be viewed as a Bayesian method. 5 Australian Bank Bill Rate: Kalman Filter Approach approximation for forecasting future long-term interest rates" although it is not suited for testing market efficiency. In this paper an AR(1)2 model based on equation (6) is estimated to provide another basis for comparison with the Kalman Filter approach: r t = α + β r t-1 + ε 2,t (6) where α and β are constant parameters similar to the time invariant model; r t-1 is the actual bill rate at time t-1; ε 2 is the usual regression error term. The forward rates used in the time invariant and the Kalman Filter approaches are computed from the Australian 90-day and 180-day bank bill rates. The method used is similar to that developed by Stigum [1981] for use with discount securities rather than coupon equivalent yields. This is referred to as the "bill parity" method and is given by: dr* = [ 1 - ( 1 - dr1 t1/365 ) / (1 - dr2 t2/365) ] * 365/(t1 - t2) (7) where, dr1 is the discount rate on the longer term to maturity instrument which has t1 days to maturity, dr2 is the discount rate for the shorter term to maturity instrument which has t2 days to maturity (t1 > t2); dr* is the forward rate for the period t2 and t1. The "bill parity" method provides an accurate estimate for discount securities such as bank bills, since it is constructed to exactly match the holding period returns of two strategies. The first is buying the longer-term bill and the second is buying the shorter-term bill and rolling the investment over into the next shortterm bill. 2Random walk is a special case of an AR(1) process. This was correctly pointed by an anonymous referee and the reference is Kendall and Ord [1990, page 57]. 6 Australian Bank Bill Rate: Kalman Filter Approach In order to compare the three approaches i.e. the time invariant parameter, AR(1) and the Kalman Filter, each of the three models is examined for general forecast accuracy with one step ahead forecasts beyond the sample period. The accuracy of forecast is measured in terms of root mean square error (RMSE), mean absolute error (MABE) and Theil's U-statistic. To establish whether there are statistically significant differences in relative predictive abilities, the Ashley, Granger and Schmalensee [1980] (AGS) method is then employed. Further explanation of the AGS test is included in Section 5 where the results are analysed. 4. DATA USED IN THE STUDY: Daily closing rates (end of day) for Australian bank accepted bills for two maturities, 90 days and 180 days were obtained from the Reserve Bank of Australia. The data cover the period January 1990 to March 1993. Figure 1 reveals the general nature of bank bill rates during this period, containing month-end figures only. The falling interest rate outlook over this entire period is also evident from the fact that the 180-day rates are typically below the 90-day rates. The "bill parity" relationship, as explained earlier, is applied to compute the 90-day rate in 90 days’ time i.e. t-kft, in terms of the notation of this paper, where k = 90. This implied forward rate is the regressor in both the time invariant and the time varying parameter models. To allow computation of the forward rate, the sample period used to estimate the model parameters begins in April 19913. In all, 520 observations between 2nd April 1990 and 24th September 1992 are utilised for parameter estimation and 100 observations thereafter are used for forecasting. predictive accuracy of the models are tested over the forecast period. The Table 2 reports the descriptive statistics of the estimation and the forecast periods for both the 90-day and 180-day bill rates. 3This is because there are no observations for the regressor, the 90-day forward rate, for the first 90 days of the data. 7 Australian Bank Bill Rate: Kalman Filter Approach 5. ANALYSIS OF THE RESULTS: 4 Table 3 provides the statistics of the time invariant parameter model. The adjusted R-square of 0.96 is sufficiently high indicating reasonable explanatory power of the relationship expressed by equation (2). The estimated parameters are also highly significant as can be seen from the tstatistics. Granger and Newbold [1974] point out that as most macro economic data are integrated, the standard significance tests in regressions involving levels of such data may be misleading. They suggest that a critical t-value of 11.25, rather than the standard normal value of 1.96, be used for significance tests. The parameter β of the time invariant model still appears significant even when the higher critical t-value is used. Furthermore, the least square estimators are unbiased and the Gauss-Markov theorem holds whether or not the regressor is stochastic (see Greene [1993], page 184). The parameter estimates of the time invariant model, however, appear to be inefficient for two reasons. First, as Table 3 confirms, the OLS residuals are serially correlated. Second, the assumption of absence of heteroscedasticity is violated. White's direct test for the presence of heteroscedasticity (White [1980]) and the Breusch-Pagan statistic (Breusch-Pagan [1979]) both support the presence of heteroscedasticity. The F-statistic in Table 3, for the joint test of the null hypothesis that α = 0 and β = 1, indicates that the hypothesis can not be accepted. This is consistent with Chiang and Kahl [1991]. Further, testing constancy of the parameter estimates with a Chow test (see Greene [1993] pp. 211-212) 4All statistical computations were made using the package TSP/386 on a i486 personal computer running at 33MHZ. 5One of the referees correctly points out that the high R-square may be spurious given the values of DW statistic and Q(10) in Table 3. The Granger and Newbold (1974) approximation may also be optimistic, particularly, due to the strong large sample property obtained by Phillips (1986). The parameter β in Table 3 remains significant even after applying the correction suggested by Phillips (1986). 8 Australian Bank Bill Rate: Kalman Filter Approach establishes that these parameters are not time invariant within the sample period6. As part of the residual diagnostic in Table 3, the Augmented Dickey-Fuller test suggests that the residual series is I(1) i.e. it has a unit root. This indicates that the structural relationship represented by the time invariant parameters in equation (2) are non-stationary. Table 4 provides statistics for the AR(1) model. This model has a much smaller sum of squared residuals than the time invariant parameter model and consequently a higher adjusted R-square value. The diagnostic statistics suggest that the residuals are largely serially uncorrelated and there is no evidence of heteroscedasticity. The t-statistic of the estimated parameter α is not statistically significant, although the joint test hypothesis that α = 0 and β = 1, is significant, implying rejection of the hypothesis. The results in Table 4 are similar to those obtained by Chiang and Kahl [1991] but offer much stronger support for the AR(1) model. Chiang and Kahl [1991] use quarterly observations rather than daily observed rates used here7. The results for the Kalman Filter approach in Table 5 show time variation of the parameters. Quarterly values of mean α t and mean β t and the corresponding standard deviations covering the entire sample period are given in this table. The final estimates of the Kalman Filter procedure for α t and β t show that these are significantly different from 0 and 1 respectively. Both these parameter series exhibit a first order serial correlation coefficient exceeding 0.99. These observations suggest that the parameters α t and β t are both time varying and the slope coefficient is generally not equal to 1. 6The actual hypothesis tested is that the parameters are the same across the mid-point of the sample. 7It may be that daily data exhibit stronger AR(1) behaviour than quarterly data. 9 Australian Bank Bill Rate: Kalman Filter Approach This result is consistent with the existence of a risk premium as suggested by Lauterbach [1989]. However, the average of α t (0.09134) is not equal to the intercept parameter of the time invariant model ( -0.0022) as suggested by Garbade and Wachtel [1978]. The average of β t (0.07028) is also not equal to the slope parameter of the time invariant model (0.9229). Furthermore, it may be argued that a piecewise OLS approach as suggested by Fama [1984] is applicable here but the time variation of the parameters observed in this study points to a continuous variation rather than step changes. The performance of the three models is compared using two approaches. The first, based on the estimation period, compares the sum of squared residuals and residual variance. The second, using a forecast period, compares the models using the root mean squared error (RMSE), mean absolute error (MABE) and Theil's U-statistic. In the estimation period comparison, the Kalman Filter approach achieves both the minimum sum of squared residuals and the minimum error variance, which represents an order of magnitude improvement over the other two models. This is followed by the random walk model. The second comparison reported in Table 6 analyses the one step ahead forecasting accuracy of the three models. In terms of the three measures, the time invariant parameter model performs poorly. The performance of the AR(1) and the Kalman Filter approach is similar. To identify any significant difference in the relative predictive abilities documented in Table 6, the Ashley, Granger and Schmalensee [1980] (AGS) procedure is employed. The AGS test is a pairwise examination of statistical difference between two respective mean squared errors (MSEs). Essentially, the test involves estimating the regression equations shown in Table 7, and the hypothesis tested is also defined in that table. Since it is a pairwise test and there are three techniques involved, Table 7 analyses all three possible combinations. Of interest, however, is the comparison between the random walk and the Kalman Filter approaches since Table 6 establishes the poor performance of the time invariant parameter model . The last panel in Table 10 Australian Bank Bill Rate: Kalman Filter Approach 7 compares these two techniques and the statistics of a3 and b3 establish the dominance of the random walk model, although the coefficient values are small. 6. CONCLUSIONS: This paper examines the relationship between forward rates and future spot rates in the Australian bank bill market with the help of time varying parameters. Of the two models which relate the implied forward rates to future spot rates, there is evidence to reject the time invariant parameter model though the Kalman Filter technique appears to capture the essence of time variation in the parameters. In the Kalman Filter approach both the intercept parameter ( α t ) and the slope parameter ( β t ) are random walk (plus noise) in nature and the measurement equation is a linear transformation of these parameters. Further, the estimation period analysis suggests rejection of the expectation hypothesis. This result also differs from the AR(1) model, where the model basically relates "tomorrow's" rate to "today's" rate. In terms of forecasting ability, both the AR(1) and the Kalman Filter approaches perform better than the time invariant parameter model. There is also some evidence that the AR(1) model is better able to forecast than the Kalman Filter approach. In this respect, the result differs from the findings of Chiang and Kahl [1991] in their study of US Treasury Bill rates. Perhaps one reason for this difference is the use of daily data whereas Chiang and Kahl use area, however, requires further investigation. 11 quarterly data. This Australian Bank Bill Rate: Kalman Filter Approach REFERENCES Ashley, R., Granger, C. W. J, Schmalensee, R., [1980], "Advertising and Aggregate Consumption: An Analysis of Causality", Econometrica, 48, 1149-67. Berger, A. N., and Craine, R., [1989], "Why Random Walk Models of the Term Structure Are Hard To Reject", Journal of Business and Economic Statistics, 7, 161-167. Breusch, T., and Pagan, A., [1979], "A Simple Test for Heteroscedasticity and Random Coefficient Variation", Econometrica, 47, 1287-1294. Carew, Edna, [1991], The Financial Markets in Australia, Allen & Unwin, North Sydney. Chiang, T. C. and Kahl, D. R., [1991], "Forecasting The Treasury Bill Rate: A Time Varying Coefficient Approach", Journal of Financial Research, 4, 327-336. Cole, S. C., Impson, M., Reichenstein, W., [1991], "Do Treasury Bill Futures Rates Satisfy Rational Expectation Properties?", Journal of Futures Markets, 11, 591-601. Fama, E. F., [1984], "The Information in the Term Structure", Journal of Financial Economics, 13, 509528. Friedman, B. M., [1980], "Survey Evidence on the 'Rationality' of Interest Rate Expectations", Journal of Monetary Economics, 7, 453-465. Garbade, K., and Wachtel, P., [1978], "Time Variation in the Relationship Between Inflation and Interest Rates", Journal of Monetary Economics, 4, 755-765. Granger, C and Newbold, P., [1974], "Spurious Regressions in Econometrics", Journal of Econometrics, 2, 111-120. Greene, W. H., [1993], Econometric Analysis, Macmillan Publishing Company, New York. Hafer, R. W., Hein. S. E., MacDonald, S. S., [1992], "Market and Survey Forecasts of the ThreeMonth Treasury Bill rate", Journal of Business, 65, 123-138. 12 Australian Bank Bill Rate: Kalman Filter Approach Hamburger, M. J. and Platt, E. N., [1975], "The Expectation Hypothesis and the Efficiency of the Treasury Bill Market", Review of Economics and Statistics, 57, 190-199. Harvey, A. C., [1990], Forecasting, Structural Time Series Models and the Kalman Filter, Cambridge University Press, Cambridge. Kahl, D. R. and Ledolter, J., [1983], "A Recursive Kalman Filter Forecasting Approach", Management Science, 29, 1325-1333. Kamara, A., [1990], "Forecasting Accuracy and Development of a Financial Market: The Treasury Bill Futures Market", Journal of Futures Markets, 10, 397-405. Kane, E. J., [1983], "Nested Tests of Alternative Term Structure Theories", Review of Economics and Statistics, 65, 115-123. Kendall, M. and Ord, J. K., [1990], Time Series, Oxford University Press, New York. Lauterbach, B., [1989], "Consumption Volatility, Production Volatility, Spot-Rate Volatility, and the Returns on Treasury Bills and Bonds", Journal of Financial Economics, 24, 155-179. MacDonald, S. S., and Hein, S. E., [1989], "Futures Rates and Forward Rates as Predictors of NearTerm Treasury Bill Rates", Journal of Futures Markets, 9, 249-262. Muth, J. F., [1961], "Rational Expectations and the Theory of Price Movements", Econometrica, July, 315-335. Phillips, P., [1986], "Understanding Spurious Regressions", Journal of Econometrics, 33, 311-340. Stigum, M., [1981], Money Market Calculations: Yields, Break-Evens, and Arbitrage, Homewood, Ill., Dow-Jones-Irwin. White, H., [1980], "A Heteroscedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroscedasticity", Econometrica, 48, 817-838. 13 Australian Bank Bill Rate: Kalman Filter Approach Table 1 Authorised Dealers - Selected Assets a Year Bills b CDs c 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 258 265 361 449 306 286 367 502 308 383 405 161 25 85 177 231 213 207 115 214 274 422 597 226 a Average of weekly figures in $ millions. Source: Lewis, M. K., and Wallace, R. H., [1993], The Australian Financial System Longman Chesire, South Melbourne, page 372. b Commercial bills accepted or endorsed by banks. c Includes some other bank securities. 14 Australian Bank Bill Rate: Kalman Filter Approach Table 2 Dataset Descriptive Statistics 90 Day Bill Rate Statistics Mean Std. Dev. Skewness Kurtosis(Excess) Maximum Minimum 180 Day Bill Rate Estimation Forecast Estimation Forecast 0.10285 0.03008 0.07637 -1.21163 0.15300 0.05450 0.05811 0.00172 -1.20196 3.59484 0.06000 0.05240 0.10172 0.03007 0.08388 -1.18760 0.15270 0.05320 0.05827 0.00218 -1.94675 2.44954 0.06130 0.05180 Estimation period : 2nd April 1990 - 24th September 1992 (520 observations) Forecast period : 28th Sept 1992 - 22nd March 1993 (100 observations) 15 Australian Bank Bill Rate: Kalman Filter Approach Table 3 Statistics of Time Invariant Expectation Modela Regression: Sample size Adjusted R-squared Sum of squared residual Residual Diagnostic DW-statistic Ljung-Box Q(10) Augmented Dickey-Fuller test Breusch-Pagan heteroscedasticity test White's heteroscedasticity test Error variance Co-efficients α β H0: α = 0, β = 1; F-test Chow-test (sample split at mid point) 520 0.96 0.01523 0.0498 3775.14 (0.000) -2.26 (0.514) 15.32 (0.000) 25.06 (0.000) 0.2941E-04 -0.0022 (0.022) 0.9229 (0.000) 1123.03 (0.000) 29.06 (0.000) a r t = α + β t-90 f t + ε1,t r t = 90 Day Rate at t, t-90 f t = 90-day forward rate at t implied by 90day and 180-day rate at t-90, ε1 = Regression residual Numbers in parentheses are p-values. 16 Australian Bank Bill Rate: Kalman Filter Approach Table 4 Statistics of AR(1) Walk Model a Regression: Sample size Adjusted R-squared Sum of squared residual Residual Diagnostic DW-statistic Ljung-Box Q(10) Augmented Dickey-Fuller test Breusch-Pagan heteroscedasticity test White's heteroscedasticity test Error variance Co-efficients α β H0: α = 0, β = 1; F-test Chow-test (sample split at mid point) 520 0.99 0.00029 1.79 12.78 (0.236) -20.48 (0.000) 0.93 (0.334) 1.55 (0.462) 0.5612E-06 -0.41E-04 (0.697) 0.9986 (0.000) 15.74 (0.000) 1.06 (0.347) a r t = α + β r t-1 + ε2, t r t = 90-day rate at t, r t-1 = 90-day rate at t-1 , ε2 = Regression residual Numbers in parentheses are p-values. 17 Australian Bank Bill Rate: Kalman Filter Approach Table 5 Statistics of Kalman Filter Approach a Sum of squared residual Error variance Mean α t Mean β t Final estimate of α t 0.0001782 0.0000003 0.09134 0.05377 0.04829 (0.000) b 0.07028 (0.001) b Final estimate of β t Time variation of α and β Period Mean α Std. Dev. α Mean β Std. Dev. β Apr. - Jun. '90 Jul. - Sep. '90 Oct. - Dec. '90 Jan. - Mar. '91 Apr. - Jun. '91 Jul. - Sep. '91 Oct. - Dec. '91 Jan. - Mar. '92 Apr. - Jun. '92 Jul. - Sep. '92 0.15604 0.14994 0.12094 0.10801 0.09934 0.09161 0.07611 0.06750 0.06244 0.05259 0.00621 0.00365 0.01046 0.00364 0.00435 0.00298 0.00542 0.00149 0.00333 0.00203 -0.03250 -0.05494 0.03588 0.07174 0.07889 0.08032 0.07905 0.09070 0.06618 0.07744 0.03691 0.01679 0.04417 0.01543 0.00484 0.00231 0.00539 0.00499 0.00757 0.00270 a r t = α t + β t t-90 f t + ε3, t r t = 90-day rate at t, t-90 f t = 90-day forward fate at t implied by 90-day and 180-day rate at t-90, ε3 = Regression residual α t and β t time varying co-efficients; b p-values for t-distribution testing parameter = 0. 18 Australian Bank Bill Rate: Kalman Filter Approach Table 6 Forecasting Accuracy a RMSEb MABEc U-Statd Time Invariant Expectation 0.00712 0.00627 0.06462 AR(1) Model 0.00038 0.00024 0.00330 Kalman Filter Model 0.00055 0.00026 0.00475 a Forecast period is 100 days after 24th September 1992; One step ahead forecast by the random walk model and Kalman Filter. Forecast by time invariant model relies on parameters obtained from the estimation period. b Root Mean Squared Error c Mean Absolute Error d Theil's U statistic 19 Australian Bank Bill Rate: Kalman Filter Approach Table 7 AGS Test Comparing Mean Squared Error (MSE)a Comparing e1 and e2 e1,t - e2,t = a1 + b1 [ (e1,t + e2,t) - (me1 + me2,)] + u1,t H0: MSE 1 > MSE 2, Requires a1 and b1 non-negative and at least one of them strictly positive a1 -0.00609 (-84.73) b1 0.93493 (47.25) Adjusted R-square 0.96 DW statistic 1.54 Comparing e1 and e3 e1,t - e3,t = a2 + b2 [ (e1,t + e3,t) - (me1 + me3 )] + u2,t H0: MSE 1 > MSE 3, Requires a2 and b2 non-negative and at least one of them strictly positive a2 -0.00629 (-63.88) b2 0.87434 (33.18) Adjusted R-square 0.92 DW statistic 1.13 Comparing e2 and e3 e2,t - e3,t = a3 + b3 [ (e2,t + e3,t) - (me2 + me3 )] + u3,t H0: MSE 2 > MSE 3, Requires a3 and b3 non-negative and at least one of them strictly positive a3 -0.00019 (-7.30) b3 -0.19829 (-6.47) Adjusted R-square 0.29 DW statistic 1.39 a Ashley, Granger, Schmalensee (1980) test for statistical difference between two respective mean squared errors. This requires that mean errors are positive for both series. If it is positive only for one series then two times the mean of the negative forecast error series is added to the dependent variable. This adjustment will allow the test to measure the mean difference in absolute forecast errors. Error defined as (Model Forecast - Market Observed) e1 error for Time Invariant Expectation Model; me1 mean for this error e2 error for AR(1) Model; me2 mean for this error e3 error for Kalman Filter Model; me3 mean for this error u1, u2, u3 regression error; t-stats are in parentheses 20 Australian Bank Bill Rate: Kalman Filter Approach Figure 1 Bank Bill Yield : Estimation & Forecast Periods a 17 bab90 16 15 14 bab180 13 12 11 10 9 8 7 6 5 Jun '90 Jun '91 Dec '90 Jun '92 Dec '91 Dec '92 a Yield is expressed as percentage per annum; Month-end figures shown BAB90 refers to 90-day bank accepted bill BAB180 refers to 180-day bank accepted bill 21