Predicting Municipal Bond Yields: Can Municipal Indices Predict Individual Bond Yields? by Le Duong University of Arkansas, Little Rock and Duane Stock University of Oklahoma 205A Adams Hall Price College of Business Norman, OK 73019 Email: dstock@ou.edu September 14, 2009 Presented at University of Oklahoma Center for Financial Studies workshop. The authors appreciate comments by seminar Jesus Salas, Louis Ederington, Chitru Fernando, and Scott Linn. 0 Abstract Recent studies have shown varying degrees of predictability in returns and volatility for different types of financial instruments. Of course any predictability should be exploited by those dealing in the instruments. To date, there has been little analysis of predictability in municipal bond yields and volatility. We develop econometric models of municipal bond indices and then test for any usefulness in predicting yields of individual bonds. The sample is restricted to larger negotiated, general obligation yields that meet strict tests of liquidity and quality of yield data. We apply different types of econometric models including ARMAGARCH and alternative logit models where, in one type of logit model, volatility helps predict the sign of yield changes. Then we develop strategies to exploit the predictability. In summary, we find there is evidence of useful yield predictability that produces beneficial strategies. The predictability and strategies are robust to different specifications. 1 Introduction Every year United States corporations and municipalities issue huge amounts of debt. As of the first quarter of 2009, there was $6,772 billion of corporate debt and $2,669 billion of municipal debt outstanding. If the issuer can save as little as a few basis points due to yield predictability, the life- time costs can be reduced by millions. As an example, Lambert and Pierog (2009) report that in the first half of 2009 alone, California sold $13.4 billion of municipal bonds. Reducing costs by 10 basis points on these first half issues would have saved $13.4 million a year. Any possible small yield predictability is particularly useful to bond issues with flexible timing. For example, negotiated municipal issues and corporate shelf issues have a good deal of timing flexibility. The purpose of our study is to test the ability of simple econometric models of municipal yield indices to predict yields of individual municipal bonds. Our results are potentially useful to anyone buying or selling municipal bonds including mutual funds, hedge funds, banks and individual investors. For convenience and quality of data, we use the context of timing an issuance. We include alternative models of municipal indices where some include both a yield index forecast and a volatility forecast whereas others may only include a yield index forecast. Our tests are strong in that we test the models on out of sample actual yields of specific bonds, at issuance and soon after, as opposed to testing the ability to merely forecast the index itself for periods before and after actual issuance. We note that forecasting a municipal bond index is very problematic due to non-synchronous trading. It is important to not confuse our analysis with a test of optimal timing for a window including dates before actual issuance. 2 More specifically, we utilize daily, weekly, and monthly index data to aid in timing (trading) for forty large, negotiated, general obligation bonds. The econometric models for timing bond issues will be helpful in two ways. First, the models will give predictions of changes (or the sign thereof) in yield of indices and individual bonds. Second, some models will simultaneously give conditional volatility of the yield changes. Analyzing both level and volatility of yields can clearly enrich analysis of the timing decision. With regard to volatility, an obvious question is how much will variance of expected yields increase if the planned issuance (transaction) date is delayed one or two periods? Furthermore, in one type of logit model, volatility modeling can help predict the sign of yield changes. Timing is of course complex. At some point(s) in time the issuer (trader) has to consider all the data available and decide to issue (trade) or not within a certain time frame. Even deciding a cutoff for when to stop and analyze available data is part of the process. To permit usage of the best possible bond specific data, we stress analysis from the vantage point of the close of the period immediately prior to actual issuance. We note that for dates prior to issuance, it is impossible for a researcher to know the precise yield at which the specific bond “could have” been issued. Yield indices are problematic as they may or may not provide precise answers given the myriad of bond specific factors affecting yield and, furthermore, non-synchronous trading is quite problematic. Our primary question is as follows. Given all the information available at the close of the day (week, month) before issuance, can an econometric model of an index predict actual individual bond yields for the issuance day and, also, immediately following periods? What alternative model(s) could the issuer have potentially used to execute superior issuance timing? Furthermore, is predictability robust across different models? Our results clearly 3 suggest that econometric models such as those we estimate can be developed to help time trading and issuance of individual municipal bonds. Our question is in contrast to an ex post issuance performance examination of a broad window both before and after issuance which answers an alternative question: Did the issuer pick the minimum index (as opposed to bond specific) yield within that window? See Kadapakkam and Kon (1989) where, using government bond indices as opposed to bond specific yields, they tested for optimal timing of corporate shelf issues as compared to nonshelf issues. Such an analysis is obviously important but develops no model for issuance timing of a specific bond, does not consider volatility of yields, and does not consider actual prices of an individual bond with numerous issuer and issue-specific characteristics. The next section describes recent evidence on the benefits of developing models to predict returns, yield changes, signs of yield changes, and volatility for different types of instruments. Then we describe our data sources where we stress that although indices are necessary to develop econometric models, precise and unbiased strategy testing issuance timing (trading) uses yields of individual bonds at issuance and immediately after. The next section estimates the models. Then we execute strategies to test the usefulness of the models for daily, weekly, and monthly periods. The conclusion summarizes the research. Predicting Returns, Yields, Signs of Yield Changes and Volatility Recent advances in conditional distributions of financial time series have greatly advanced modeling of expected returns of equities and bonds, yield changes and conditional variance. Jones, Lamont, and Lumsdaine (1998) and, also, Baker, Greenwood, and Wurgler (2003) suggest that yields are predictable. Cochrane (1999) lists return predictability as one 4 of the new facts of finance and maintains that bond yields and returns are predictable. 1 Papageorgiou and Skinner (2002) provide evidence of government bond yield predictability. Cochrane and Piazzesi (2005) find that excess returns of government bonds are predictable. Egorov, Hong, and Li (2006) suggest affine term structure models have potential for predicting yields.2 Christoffersen, and Diebold (2006) note that researchers, such as White (2000) and Pesaran and Timmerman (2004) have been successful in predicting the sign, if not the magnitude, of asset returns.3 Green, Li, and Schurhoff (2008) find that lagged municipal price changes are correlated with municipal yields. Also, conditional variance (volatility) of returns for many types of securities has been clearly proven predictable. See Andersen, Bollerslev, Christoffersen, and Diebold (2006) for a survey of volatility forecasting. Given the extensive testing of predictability for other instruments, a test and related application for the municipal bond market seems overdue. Numerous authors have shown that even a small amount of predictability of returns and volatility, even if not clearly statistically significant, can have a large impact on financial strategy and decisions. Among others, Kandel and Stambaugh (1996) have shown that weak return predictability can have a strong impact on portfolio decisions. Xu (2004) finds that small levels of equity return predictability can be utilized to develop strategies with both higher returns and less risk. Furthermore, Xu (2004) stresses that it is not the statistical significance of return and volatility estimations that matter so much as does the profitability 1 We note that Cochrane (1999) and some others noted may tend to analyze data with lower frequency than ours. 2 Some suggest that predictability of financial time series is due to time varying risk premia. 3 There is also considerable evidence that equity returns are predictable. For one recent example, see Cohen and Frazzini (2008). 5 of any weak statistical significance. Barberis (2000) finds that small levels of return predictability affect asset allocation. Some may suggest that the ability to predict returns is in violation of market efficiency but Marquering and Verbeek (2004) and Lewellen and Shanken (2002) maintain this is not necessarily the case. In this context, rational asset pricing implies time varying risk premia that lead to predictable patterns. Our models of expected yields (yield changes) translate into useful strategies that are an improvement over naïve, random timing for at least two reasons. First, the above cited studies on equities, bonds, and other instruments suggest that any weak predictability given by estimated mean and volatility equations leads to superior performance. Second, our application is one where frequent transactions typically dictated by predictability and related dynamic strategies will not occur. In this context, it has been known for decades that, for example, equities sometimes exhibit small autocorrelation in returns. See, for example, Neiderhoffer and Osborne (1966), Fama (1970, 1991), Conrad and Kaul (1989) and Lo and MacKinlay (1999). More recent research suggests market microstructure may generate autocorrelation. Nonetheless, investment strategies developed to exploit any autocorrelation frequently find that any predictable autocorrelation in returns is frequently more than offset by transactions costs needed to make numerous trades. Our research problem is one where the issuer chooses to issue the bonds at only one certain time. Of course, issuance (transactions) costs thus occur only once where the level of costs are unaffected by the day chosen to issue the bonds.4 Therefore, there is a greater likelihood that exploitation of any 4 Kadapakkam and Kon (1989) find that, ex post, corporate bond shelf issues seem to be better timed than nonshelf. Adding on to our earlier comments, we note that their work does not analyze municipal bonds nor does it report models of timing or volatility. 6 weak predictability in expected yield changes and volatility will lead to beneficial timing strategies. With regard to volatility, Fleming, Kirby, and Ostdiek (2001 and 2003) and Gomes (2007), among others, predict volatility and demonstrate the economic gains from doing so. For our application, we note that volatility modeling can be useful in more than one way. Consider how volatility affects confidence of yield predictions. That is, if the model predicts a decline in yields, thus encouraging a delay in issuance, it may be especially appealing to delay if the expected decline is large and, also, the volatility is small as the issuer can be relatively confident of a decline in yields. On the other hand, if the model predicts large increases in yields and all the expected volatilities are small, it is very unappealing to delay issuance as the likelihood of greater yields seems strong.5 In order to enhance our volatility analysis and also develop a broad sample representing various environments, our sample of bond issuances is developed from four distinctly different periods where yields tended to have different trends (upward and downward) and levels (high and low) of volatility. Volatility modeling can also be useful if it can forecast the level of yields. It is logical to suggest that predictable volatility could lead to predictable risk premia, returns, and yields. Thus, GARCH in the mean models may offer potential. More central to our research, Chrisoffersen and Diebold (2006) show how predictability in the sign of returns and volatility may well be interrelated. Any slight sign predictability can potentially be very useful for bond issuers constantly asking themselves the question: Will the interest cost of 5 Also, we of course note that if a large increase in rates is predicted and expected volatility is high, there is a strong likelihood rates will increase very dramatically. 7 the bond issue rise or fall if issuance is delayed one day, two days, one week, two weeks, one month, etc. After lengthy testing of alternative specifications and explanatory variables to best answer our questions, we report the parsimonious models that worked best. The base model is the ARMA-GARCH which predicts both changes in level of yields and volatility. In the context of sign prediction only, we also report a simple logit model for sign prediction (ARMA-Logit) that has no volatility component. Furthermore, we report an alternative logit model (CD-Logit)6 suggested by Christoffersen and Diebold (2006) which uses volatility to predict signs of yield changes. Alternative models are used to test robustness of any claims of predictability. Data Sources and Sample Selection We are aware of controversies in claiming predictability in, for example, equity returns. In one of the earlier studies, Lo and MacKinlay (1992) find equity predictability is strong and economically significant. In more recent examples, Cochrane (2007) defends return predictability in equities whereas Ang and Bekaert (2006) find that long run return predictability is not significant. It is important that Ferson, Sarkissian, and Simin (2003) describe how spurious regressions and data mining can lead to spurious predictability. Thus we develop our sample and construct our tests to be as free as possible of bias. 6 We describe two alternative CD-Logit models later. 8 The dominant principles of our sample selection are 1.) quality of individual bond yields and 2.) availability of indices that match characteristics of these specific bonds. Thus, we match individual general obligation (GO) bonds and indices of GO bonds described in more detail below. Individual General Obligation (GO) Bond Yields We examine forty large-issuance, negotiated, general obligation (GO), noncallable, municipal bond issuances, given in Table 1, for ex post optimal issuance timing. We choose negotiated issues as there is much more issuance flexibility than for competitive issues and we can thus more credibly analyze the actual chosen issuance timing where the same can be said of corporate shelf issues versus nonshelf issues. We only use bonds where the Bloomberg financial system provides an issuance yield (yhp,issue) and a closing yield for every trading day of the following four weeks. This requirement assures high quality of data and also allows monthly analysis. Bloomberg issuance yield is not necessarily the closing yield for the day but the yield at time of day when issued. We use a type of bond that is as homogeneous as possible such that an index (described below) based on such bonds is a credible average representation of the individual bond issues. In this context we note that Harris and Piwowar (2006) find that complex municipal bonds have higher costs than simple bonds. General obligation bonds are known for frequently being typically less complex and more homogenous in sources of debt service than revenue bonds. Noncallable bond issues are more homogenous in that the value of call features can vary substantially due to variation in call price, call protection period, etc. even when matching maturity and credit quality. To appreciate the importance that callability has 9 upon yields we refer readers to Duffee (1998) who stressed it is very problematic that many corporate bond indices mix both callable and noncallable bonds. His empirical research found that a sample of callable bonds had very different sensitivity to changes in the level of interest rates than an otherwise similar sample of noncallable bonds. We choose bonds with high credit quality as we think they are more homogenous than lower credit quality. Although the above restrictions severely restrict sample size, we feel the need for such requirements to assure accuracy in yields and strategic testing, especially since municipal bonds are well known for illiquidity. As mentioned, our source for bond specific yields was the Bloomberg financial information system. To be sure we proportionately include issues from various types of interest rate environments, we classified the largest negotiated, GO, noncallable municipal bond issuances in the period January 2000 to August 2006 available on Bloomberg into the following continuous, non-overlapping four groups. (1) Issued between January 2000 and October 2001. This was a period when municipal bond interest rates were generally decreasing and rate changes exhibited relatively low volatility. (2) Issued between November 2001 and July 2003. This was a period when municipal bond interest rates were generally decreasing and rate changes exhibited relatively high volatility. 10 (3) Issued between August 2003 and July 2005. This is a period when municipal bond interest rates were generally increasing and rate changes exhibited high volatility. (4) Issued between August 2005 and September 2006. This was a period when municipal bond interest rates were generally increasing and rate changes exhibited relatively low volatility. The above categories identify contrasting environments and also allow analysis of how timing may be more important, for example, in a more volatile environment than a less volatile environment. Then from each of the four groups above, we sorted municipal bond issuance into two credit quality groups, i.e., AAA and AA. From each of the groups, we chose the five largest bond issuances that also had closing yields available for every trading day in the four weeks after issuance to allow monthly analysis. We chose the largest because one may argue timing is more critical the larger the issue and larger issues tend to be more liquid. The earliest issuance used was January 24, 2000 and the latest was August 14, 2006. The Bloomberg system calls their yields individual HP (historical price) closing yields and reports an average of yields across a sample of brokers and dealers (“price contributors”) on new issues. Where available, yields are reported on a daily basis after issuance. For examination of timing strategies, we feel post- issuance HP yields provide superior information compared to using index yields. Studies of bond issuance using index data to determine timing performance are not as precise as we wish for individual bonds. However, when we analyze the timing of a particular issue, HP yields on individual bonds are not available before issuance. Therefore we need some measure of yields before 11 issuance in order to develop predictions and strategies for timing the issuance of a specific individual bond issue. Hopefully the particular index chosen will represent the individual bonds well enough to estimate models accurately enough to credibly suggest usefulness. Thus, the second type of yield data was a Bloomberg daily time series index of noncallable general obligation bonds with the same maturity and credit rating as the particular issue. The authors were fortunate Bloomberg provides such a good match to our preferred bond features. Alternative indices that we found were not as good a match. For example, consider the S &P National Municipal Bond Index where the maturity requirement is only that bonds in the index have a maturity of more than one month where there is thus a large variation in maturity. Furthermore, callable bonds are mixed with noncallable bonds and the credit rating is anything BBB- or above. Using the better Bloomberg index data, we estimate expected yield changes and volatility using data through and including closing yields the day (week, month) before issuance (t-1). As for individual HP yields, the Bloomberg index yields are derived from market quotes, not matrix prices. In order to construct their yield series, Bloomberg collects the daily closing of bid and ask quotes from reliable dealers. For each maturity and credit quality, for example, one year AAA, they average the bid and ask quotes of their sample bonds to derive a given index yield. The Bloomberg index data begins on January 2, 1992 for AAA bonds and on June 6, 1994 for AA bonds where we use index yields through September 2006. Using simple strategic decision rules, we evaluate the actual issuance for timing. That is, was it wise to issue on the date chosen or would it have been better to defer issuance? If our strategies show good performance, then the index models estimated are useful for future applications. Any issuer 12 can easily imagine a point in time is t-1, estimate a model for future values of the index, and decide the timing of an issuance (trade). We focus on the window from the period (day, week, month) just before issuance to immediate periods after issuance for the below reasons. First, it is not our purpose to forecast indices. But, of course, we need indices to forecast future yields of individual bonds not yet in existence. The cleanest and strongest test of our econometric models is the ability to forecast issuance yields of individual bonds which can be quite heterogeneous. Different studies have found the municipal market is quite heterogeneous where yields for individual bonds depend upon numerous bond specific factors. The size of issue, relative supply of municipal bonds within the state, relative demand for bonds within the state, bank pledging requirements, state income taxation, and bank qualification may all affect yields. See Kidwell, Koch, and Stock (1984) and Forbes and Leonard (1984). Furthermore, Harris and (2006) and Hong and Warga (2004) find liquidity effects on yields are dependent upon such things as credit risk and bond complexity, in addition to finding that costs are clearly higher (lower) on complex (simple) bonds. Thus, using an index to forecast yields on bonds with differing complexity is problematic. However, our sample selection and (noncallable, equal maturity, equal credit quality) index match was made to minimize such problems. Of course, it is impossible to know the yield at which a specific bond would have been issued in periods prior to issuance. Our tests could be construed as strong ‘out of sample’ tests of the econometric models of municipal bond indices. In this context, a basic question is whether the indices we have chosen are robust and well enough matched to the individual bonds so as to successfully forecast yields of the individual bonds. 13 The second reason for focusing on time periods after issuance is that predictions using index data before issuance are frequently plagued by non-synchronous (stale) trading data which leads to spurious autocorrelation in yield changes and misleading predictability of actual yields. The problems presented by non-synchronous trading are potentially quite large for municipal indices given the low liquidity of municipal bonds. Consider making decisions for, say, a window including two periods before actual issuance in period t. A regression of yield changes for the appropriate index through t-2 would be estimated and closing index yields for t-1, t, t+1, etc. predicted. However, Campbell, Lo and MacKinlay (1997) illustrate that non-synchronous trading can create a false impression of predictability in equity returns and price changes even when prices are statistically independent. As one an example of the difficulty of using data with non-synchronous trading problems, see Ahn, Boudoukh, and Richardson (2002) where they find that non-synchronous prices in a stock index create autocorrelations that are quite problematic when comparing spot and futures prices. Also, Kadlec and Patterson (1999) find that non-synchronous trading often explains much of the autocorrelation in equity returns. More recently, Schotman and Zalewsaka (2006) and Bernhardt and Davies (2008) show that non-synchronous trading has a sizeable effect. Non-synchronous trading problems are likely even larger for municipal bonds than for equities. Using index yields to predict individual HP yields is not subject to non-synchronous problems. 14 Estimations of Expected Change in Yield and Conditional Volatility ARMA-GARCH Estimations as the Base Model Daily Periods We test robustness of predictability across alternative classes of models without necessarily declaring a winner as our purpose is more to test effectiveness and then general robustness across classes of models as opposed to declaration of the best. For each class of model, we report the specifications that show the best results. Perhaps the most obvious for our analysis is the ARMA-GARCH model which simultaneously predicts change in yields and volatility. Later we do an ARMA- Logit estimation, where there is an equation to estimate (only) the sign of yield changes. Finally, we also report alternative sign prediction estimations (CD-Logit and CD-EWMA) as suggested by Christoffersen and Diebold (2006) where the sign of yield change is solely a function of volatility. We stress the prediction of sign changes as opposed to forecasts of expected (mean) changes in yields. To forecast the complete distribution of expected price (yield) changes may be too ambitious and is likely more than what is needed to be successful in timing transactions. Christoffersen and Diebold (2006) stress that profitable forecasting is more a result of forecasting market direction (sign changes) than it is forecasting mean (expected) returns. In fact, they maintain that forecasting the sign of returns (yields) can be done with surprising success and provide numerous studies to support their statement. Using index data for an ARMA-GARCH estimation, we predict changes in yield (Δyt,M) and volatility at time “t” for bonds of maturity M where Δyt,M = yt,M - yt-1,M . Our ARMA mean equation estimation uses one autoregressive (AR) term and one moving average (MA) term. In other specifications tested we employed a long list of additional 15 explanatory variables such as lagged level of yields, various measures of the slope of the term structure, various forward rates implied by the term structure, etc. but none of these additional variables were consistently significant. Furthermore, GARCH in the mean models were not helpful. Thus, we use the following parsimonious model for Δyt,M where ε t,M represents the error term of the regression. ∆yt,M = a0,M + a 1,M Δyt-1,M + b1, M εt-1,M + εt,M t2,M M 1,M t21 ,M M t21 ,M (1) (2) The following omits M from yt notation without loss of meaning. For volatility ( t ,M ) estimation at time “t” we use a standard GARCH above where 2 volatility depends only on the most recent squared residual and lagged volatility. As for the mean equation, we tested many more volatility specifications that included, for example, the lagged level of yields, forward rates, slope of term structure, asymmetric volatility, etc. but none of these additional variables were consistently significant. Strategies to Test Usefulness of Models Assume an investment banker is considering the best time to sell a negotiated, noncallable, general obligation issue. Day (week,month) t-1 has ended (or it is very early on day/week/month t) and perhaps because of other bond market activities or a myriad of other reasons, it is important to decide whether to issue as soon as possible, i.e. the next period, or, alternatively, the following period. As empirical results suggest, trends in yield changes tend 16 to last for three or more periods and our choice of issuance period depends on detecting a trend that lasts for the two periods of the issuance window. In our system, the actual day of issuance is day t. The above econometric model has been estimated using index yields through close of day (week, month) t-1 and issuers thus have index predictions for close of day (week, month) of future periods. Our strategies take advantage of predicted index yield changes for the close of these future periods. Hopefully these predicted yield changes for the index will give an indication of yields for the individual bond issues of our sample at future times and thus confirm the usefulness of the index model estimation and related strategies.7 We now introduce different strategies to test the usefulness of estimated index models for predicting individual bond yields on the day (week, month) of issuance and days (weeks, months) thereafter. Numerous rules for a strategy can be suggested. For ease of initial illustration, let us assume the choice is simplified and limited to only two days (weeks,months) and call this strategy ARMA-GARCH-A. The issuer only knows close of index for t-1 and a prediction for the index on day t. See Figure 1 for illustration of index versus individual HP yields. A simple strategy would take advantage of any trend in yields suggested by estimated index regressions. More specifically, regression results reported below show that the signs of yield changes are typically the same for each of the three periods after t-1. That is, a predicted increase (decrease) in yields from t-1 to close of t tends continues for the next period. 7 Green, Hollifield, and Schurhoff (2007) suggest that an investment banker can buy in periods before delivery and sell pieces forward ‘when issued’. 17 Thus our simple strategy is totally dependent upon the prediction for the index close of day t. Specifically, if index yields are predicted to be lower by the close of day t, delay issuance to t+1 to take advantage of the expected decline. Our estimate of the strategic issuance yield in this case is the realized Bloomberg HP yield for t+1 which is yhp,t+1. The gain (cost reduction) of the strategy is yhp,issue – yhp,t+1 where yhp,issue is the actual issuance yield for day t reported by Bloomberg. The effectiveness of this strategy hinges on the hope that the likelihood of a sizeable increase in yield on t+1 that is of greater magnitude than the expected decline on t is small. Again, expected yield changes tend to be the same sign. Robustness to a sizeable t+1 yield increase is determined by the success of execution strategies given below. Alternatively, if index yields are predicted to be the same or higher on day t, do not delay but issue on day t which is, in fact, the actual day of issuance. It is important to note that Bloomberg HP issuance yield is not necessarily the closing yield for the day but the yield at time of day when issued. Bloomberg does not report a closing yield for the individual bond on day of issuance, t, but only the issuance yield (yhp,issue) which occurs sometime between the open and close of day t. The idea is that hopefully the bond can be issued before any increase in yield occurring before the close of the day. If the strategy chooses day t, our gain calculation is zero as the strategy selects the same day (t) as the actual issuance. Note that if yields rise early on day t, the effectiveness of this strategy may be compromised. Robustness to an early day t increase in yield is determined by any success of strategy executions given below. We emphasize that all predictions of yield changes are necessarily done from the index regression model. All issuance strategy yields use the superior realized bond specific 18 HP yields. Realized individual HP yields are a stronger test of the models and strategies and are also a superior out of sample measure reflecting all unique aspects of the issue and issuer while also being free of non-synchronous trading bias. An alternative strategy to that given above is to choose to issue on the day with the lowest predicted index yield. The results of such a strategy are practically identical to that above given the persistent pattern of yield changes. We use the strategy above so that clearer comparisons can be made between ARMA-GARCH and models which only predict the sign (not magnitude) of yield changes. Alternatively, a more complex longer term strategy could be used where we call this strategy ARMA-GARCH-B. Here, as of close t-1, predictions are made for the index at close t and t+1. Issue on day t+2 if expected change in index yield is negative on both day t and t+1. The gain (cost reduction) is computed as yhp,issue – yhp,t+2 where yhp,t+2 , the individual HP yield at t+2, is the strategy yield. Issue on day t if the expected index yield change at close of day t is positive or zero and the expected index yield change at close of day t+1 is positive or zero. That is, issue the bonds before the expected increase in index yields where index yields are expected to rise on both subsequent days. The gain is zero as the strategy issuance day is the same as actual issuance day. The combination of yield changes for close t and t+1 may have mixed sign in some cases. Here we recognize that the variance around a prediction approximately doubles when the horizon is increased from one day to two days. Thus we proceed as below. 19 Defer issuance until day t+1 if the expected index yield change at the close of day t is negative and the expected index yield change at close of day t+1 is positive or zero. The gain is computed as yhp,issue – yhp,t+1 where yhp,t+1, the HP yield at t+1, is the strategy yield. If the series of expected index yield changes is positive on day t and negative on day t+1, we merely issue on day t (actual issuance) because there is no clear advantage due to the forecasts.8 The gain is thus zero because the strategy day chosen is the actual issuance day. We note that A shortcoming of the above strategies is that they are mostly based upon expected yield change and pay little attention to the expected volatility of interest rates where the greater the expected variance ( t2 ,M ), the greater the realized yield could differ from the 2 expected. Of course, a conventional risk aversion view is that the greater t ,M , the greater the possibility that rates could increase significantly if delayed (even given an expected decline) in which case delaying the issuance would be very costly. Thus the rule could be to delay if E [ Δyt+k ] < 0 , and, furthermore, E yt k Var (yt k ) 1/2 z 8 Cases where the net of positive and negative gains is negative are rare and the net is small. We chose to do this because of the volatility risk from delaying is greater as number of periods increase. 20 where k = 0, 1,…. n. That is, an issuer may require that the absolute value of the ratio of expected yields changes to variance has to meet a threshold, z. For example, greater expected decreases (increases) accompanied by low expected volatility make it more appealing (unappealing) to defer issuance.9 Our work produces a longer menu of choices than to merely issue on one of two days. Issuance could also be, for example, a week or a month hence. Thus, we also provide weekly and monthly issuance choices for those with longer planning horizons where the same logic applies. The most theoretically appealing decision may be to defer issuance until the future time when the ratio z is greatest across all given alternative horizons.10 Model Estimations and Strategy Results For each AAA-rated bond, we estimate the equations, reported in Table 2, using index yields from January 2, 1992 (origination) until the day before (t-1) the bond specific issuance date. Similarly, for each (lower rated) AA-rated bond, we estimate equations using data from June 6, 1994 to the day before the issuance date (t-1). 11 The forecasted levels and variances of yield change from t to t+2 are given in Table 3 where the range of forecasted change in (index) yield for close day t is 0.0500 to -0.0200. Successive forecasted changes in yield are 9 This is very similar to the Sharpe performance measure where here the performance is ex ante. Greater predictability represented by large expected declines in yield and low volatility suggest strong ex ante performance. 10 We assume that the issuer really needs to sell the bonds within a certain period of time. Eventually, the issuer has to sell and thus may set a threshold to sell when yields meet a certain level. 11 . The residual and squared residual results are white noise for all 40 regression results. 21 almost always the same sign but are not always so and the magnitude of the change (for the same bond) can vary considerably. The last column reports forecasted variance where the maximum one day forward variance is 0.0014 and the minimum is much smaller at 0.0002. Thus the difference in volatility among different grades of bonds and epochs in our sample can be very large. Generally, for a specific bond, there tends to be little change in variance from t to t+2. It is important to note that, given the variance for each forward period is very similar, the risk of delaying more than one period is roughly a simple multiple of the number of days delayed. 12 Of course, this means the risk adjusted expected benefit of delaying is much diminished if successive expected yield declines are small. With the above estimations, we can test the usefulness of the estimated models and the subsequent strategies. Consider Table 4. Here the decision is to either issue on day t, the actual issuance date, or t+1. If we follow this strategy, we would not issue on the actual issuance day (t) in 11 of the 40 cases. The actual dollar gain/loss from following this timing strategy is also estimated in Table 4. The gain on par value (100) from delaying until t+1 in selected cases for a specific bond issue is computed as gain = (yhp,issue – yhp,t+1) (Di) (Pi) where Di is duration of bond i and Pi is issue price relative to par.13 Of the 11 bonds where t+1 is chosen, 8 show a gain for delaying, two show a loss, and one shows no gain. This suggests the strategy is usually profitable. 12 See Andersen, Bollerlev, Christoffersen, and Diebold (2006). 13 For example, Pi is 100.5 if the bond sells at 100.5% of par. 22 The last column is the gain adjusted for the potential risk to waiting. The computation is gadj gain t Pi where rag is risk adjusted gain and g is algebraic gain per $100 par. This is analogous to a crude Sharpe measure. Now consider the more complex strategy of ARMA-GARCH- B in Table 5. If this strategy is used, issuance does not occur on the actual issuance day (t) in 29 cases. This greater disagreement with actual issuance day is to be expected given the larger menu of choices. Of these 30 cases, the gain is positive in 14 cases, negative in 7 cases, and zero in 9 cases again suggesting the strategy can be beneficial. ARMA-Logit Sign Estimation and Strategy As our second class of models, we consider logit as an alternative econometric technique to predict only the direction (sign) and not the level of index yield changes. The logit model specification for the probability (Pr) of an increase in index yields is Pr( yi 1 x i , ) e xi' / (1 e xi' ) where the dependent variable, yi, is now the probability of an increase in index yield at close of day t compared to t-1, the xi’s are the independent variables in the mean equation in Table 2, and β’s are the coefficient estimates of the same table. Of course, it is more appealing to delay issuance if the probability of an increase (decrease) is low (high). This model estimate is a probabilistic statement on the direction of index yield change where one could formulate 23 a strategy based on the strength of the probability. For example, one may consider a delay in issuance only if the probability of an increase (decrease) is less (more) than, for example, 50%. Or, one can be more demanding with the strategy and delay only if the probability of an increase is less than 40%. A weakness of this alternative (logit) procedure is that, in contrast to a GARCH process, there is no simultaneous estimation of volatility and we could find no econometric procedure for doing so. The results of Table 6 clearly suggest the strategy is useful. As in earlier strategies, at close t-1, forecasts are made for sign change of the index close at t, t+1, and t+2 where forecasts for t+1 necessarily utilize a forecast for t and forecasts for t+2 utilize forecasts for both t and t+1. This table presents the gains from the following strategy. Defer issuance until t+2, if the probability of an increase for both t+1 and t+2 is less than 0.50. Issue day t if the estimated probability of an increase in index yield for both day t+1 and t+2, is more than 0.50. Issue on t+1 if the probabilities of an increase in index yield for t+1 is less than 0.50 and the probability an increase for t +2 is more than 0.50.14 As before, the gain is calculated from the difference between yhp,issue and the HP yield (y hp, t+1 or y hp , t+2 ) chosen by the strategy. Specifically, the strategy advocates a delay of issuance until after day t in 30 cases where there is a positive gain in 17 cases, a negative gain in 6 cases, and zero gain in 7 cases. Table 7 differs from Table 6 only in that we use a 0.60 threshold for a decline instead of 0.50. The results are 19 gains, 8 losses, and 2 zero gains again suggesting the strategy is beneficial. 14 This strategy is chosen as the risk due to volatility increases with delay. 24 Christoffersen and Diebold Sign Estimation and Strategy: EWMA and GARCH Our third class of models is that given by Christoffersen and Diebold (2006) who maintain that it is possible to have sign dependence even without mean dependence in the ARMA-GARCH context given in the first class of models above. Importantly, it is volatility dependence that produces sign dependence. Furthermore, even if sign predictability occurs in concert with mean dependence, CD (2006) suggest that volatility dependence is the dominant factor in sign prediction of gains (returns). Briefly, an increase (decrease) in volatility may reduce (increase) the probability of a positive sign. As CD suggest, a logit model approximates the distribution of μ/σt. Let ∆yt be the index yield change on day t and then define the “positive index yield change” indicator as It=1 if ∆yt > 0 and It=0 otherwise. It is forecasted using a model of the form It F t et where F( ) is a monotone function with a left limit of zero and a right limit of one, µ is the expected index yield change, and σt is a forecast of yield change volatility. F( ) is determined by F ( x) exp( x) 1 exp( x) which produces the logistic regression (logit) model. We refer the reader to CD (2006) for further details on this type of modeling. 25 The μ estimate is the mean change in index yield for the previous 20 periods. 15 This roughly estimates any possible trend in yield suggested by index behavior.16 Green, Li and Schurhoff (2008) suggest there are trends in municipal yields. Then, σt is estimated as an exponential weighted moving average (EWMA) as popularized by RISKMETRICS and used by CD. We label this model CD- EWMA where the strategy is based on the probability of an increase in index yields on close of day t being less than 0.50. That is, if the probability of an index yield increase for close day t is less than 0.50, issue on day t+1, after the expected decrease in index yields at close of day t. Again, yields typically trend the same way for a number of periods after t. Otherwise, issue on day t, the actual day of issuance. Gains are calculated as before. Table 8 shows the results where, if not issued on day t, there are 12 gains and 9 losses. The CD-EWMA methodology does not readily provide for a longer window. CD (2006) also suggest that volatility could also be estimated as a GARCH process where this model is labeled CD- GARCH.17 Table 9 shows gains in 10 cases and losses in 9 for t+1 issuance. 15 Of course, bond returns are functions of yield changes. Numerous interest rate models such as Cox, Ingersoll and Ross (1985), Vasicek (1977), and Hull and White (1990) suggest interest rates may have trends due to mean reversion processes. Mean reversion leads to mean dependence. 16 Any trend in the index could be due to things such as mean reversion, sluggishness in yield changes as in Green, Li and Schurhoff (2008), and non-synchronous trading. 17 To estimate volatility, the mean equation is the ARMA (1,1) process used before. 26 Weekly and Monthly Estimations and Strategies The horizon chosen for predicting index yields can of course be longer than a two or three days. For example, the initial stages of planning a bond issue for a capital project that is not urgent may well consider alternative weeks or months. It is noteworthy that CD theory suggests results may be better for more intermediate weekly and monthly horizons.18 Thus, we also consider weekly and monthly horizons. In these cases, the time series data is less frequent and do not permit a credible GARCH estimation.19 The choice for issuance in weekly and monthly periods is parallel to daily periods. An analyst observes the close of the index at the end of week (month) t-1 which is the period before actual issuance. All the analyst knows is the closing index yield for t-1. However, they do have predictions for close of future weeks (months).The prediction for closing index value on week (month) t is from a model of index values. If the predicted index close is higher or the same at end of week (month) t, issue on actual date (day) of issuance (at the actual issuance yield, yhp,issue.) Hopefully this issuance occurs before much of the predicted rise (for the week after t-1) occurs. Alternatively, if the predicted index close is lower for week t, delay issuance to take advantage of expected lower yields. The length of the period is problematic for weekly and monthly strategies because, for example, an actual issuance could be early or late in the period after close of t-1. Any strategic issuance not in period t (actual issuance period) is at the earliest possible daily close in the week (month) t+1. CD (2006) refer to this as the sweet spot where the value of µ/σ is about 1.4. CD (2006) explain that if μ/σ is close to zero, the probability cannot deviate much from 0.5. For large μ/σ, the probability cannot deviate much from 1.0. Intermediate μ/σ have potential for much more sensitivity. 18 19 GARCH processes need large sample sizes because of the simultaneous likelihood estimation of two equations. 27 As an example, assume end of t-1 is Friday, May 14. If t is strategically chosen, actual issuance could be any day of the next week. Parallel to daily strategies, if the strategy chooses t, the gain is zero. If the strategy chooses t+1, then issuance is chosen to be the close of Monday May 24, 10 days later. If yields have risen between May 14 and May 24, the strategy will not work well. Any success in our test results implies robustness to any assumptions of this strategy. Table 10 A gives results for weekly periods. Here CD-GARCH where there are 8 gains and 9 losses whereas Table 10 B,CD-EWMA, gives 10 gains and 10 losses for weekly periods. Although not shown in a table, weekly results for an ARMA-Logit are better. If a 0.5 threshold is used, there are 16 gains and 4 losses and if a 0.7 threshold is used, there are 19 gains and 4 losses. Table 11 gives CD-EWMA monthly period results of 15 gains and 7 losses where the threshold is a 0.5 probability. Although not shown in a table, an alternative threshold of 0.7 gives 18 gains and 6 losses. Conclusion There have been strong advances concerning the predictability of returns and volatility in recent years. Such research has been conducted on equities and different types of fixed income instruments. However, there has been little applied predictability research for municipal bonds. As one example of the potential, given that the huge dollar amounts of many issues, saving just a few basis points from better issuance timing can result in huge savings. Our purpose is to test alternative econometric models for any ability to forecast and time municipal bonds issuances or other municipal bond trading. We include models that 28 simultaneously predict both yields and volatility, models that predict yields (signs of changes) only, and models that use volatility to predict yields (signs of changes). We place strong restrictions on our sample to assure data quality. Only bonds that tend to be very homogenous and demonstrate high liquidity are included. These bonds are matched with an index that matches on both credit quality and maturity. Our results suggest that there is predictability in the municipal bond market. Of course any predictability should be tested for usefulness and we thus suggest strategies for bond issuers. These strategies prove to be successful in that gains outnumber losses. This usefulness seem robust to many alternative specifications. 29 References Ahn, D.;J. Boudoukh, J. ; M. Richardson; and R. F. Whitelaw, 2002, Partial Adjustment or Stale Prices? Implications from Stock Index and Futures Return Autocorrelation”, Review of Financial Studies,15,655-689. Ang, A. and G. Bekaert, 2007, “ Stock Return Predictability: Is It There?”, Review of Financial Studies, 20, 651-707. Andersen, T. G.; T. Bollerslev; P. F. Christoffersen, and F. X. Diebold, 2006 , Handbook of Economic Forecasting, North Holland Press, Amsterdam. Cox, J.C., J.E. Ingersoll and S.A. Ross,1985, "A Theory of the Term Structure of Interest Rates", Econometrica, 53: 385–407. Baker, M., Greenwood, R. and Wurgler, J., 2003, “The Maturity of Debt Issues and Predictable Variation in Bond Returns” Journal of Financial Economics, 70, 261-291. Barberis, N., 2000, “Investing for the Long Run When Returns are Predictable” Journal of Finance, 55, 225-264. Bernhardt, D. and R. J. Davies, 2008, “The Impact of Nonsynchronous Trading on Differences in Portfolio Cross-autocorrelations”, working paper, University of Illinois (Champaign- Urbana). Cochrane J.H., 1999, “New Facts in Finance”, Economic Perspectives, Federal Reserve Bank of Chicago, Quarter 3, 59-78. Cochrane, J. H. and M. Piazzesi, 2005, “ Bond Risk Premia”, American Economic Review, 95, 138-160. Cochrane, J. , 2008, “ The Dog That Did Not Bark: A Defense of Return Predictability”, Review of Financial Studies, 21,1533-1575. Cohen, L. and A. Frazzini, 2008, “Economic Links and Predictable Returns”, Journal of Finance, 63, 1997-2011. Conrad, J. and G. Kaul, 1989, “Mean Reversion in Short Horizon Expected Returns”, Review of Financial Studies, 2, 225-240. Egorov, A. V. ;Y. Hong; and H. Li , 2006, “Validating Forecasts of the Joint Probability of Bond Yields: Can Affine Models Beat Random Walk?” Journal of Econometrics 135, 255-284. 30 Fama, E. F. , 1970, Work”, 25, 383-417. “Efficient Capital Market: A Review of Theory and Empirical Fama, E. F. ,1991, “Efficient Capital Market: II”, Journal of Finance, 46, 1575-1617. Ferson, W.E; S. Sarkissian; and T. Simin, 2003, “ Is Stock Return Predictability Spurious? “ Journal of Investment Management, 1, 1-10. Fleming, J., Kirby, C. and Ostdiek, B., 2001, “The Economic Value of Volatility Timing”, The Journal of Finance, 56, 329-352. Fleming, J., Kirby, C. and Ostdiek, B., 2003, “The economic value of volatility timing using “realized” volatility”, Journal of Financial Economics, 67, 473-509 . Gomes, F. J. , 2007, “Exploiting Short-Run Predictability”, Journal of Banking and Finance, 31,1427-1440. Green, R. C.; D. Li and N. Schurhoff, 2008, “ Price Discovery in Illiquid Markets: Do Financial Asset Prices Rise Faster Than They Fall?” , working paper, Tepper School of Business, Carnegie Mellon. Green, R. C.; D. Li and N. Schurhoff, 2007, “Dealer Intermediation and Price Behavior in the Aftermarket for New Bond Issues”, Journal of Financial Economics, 86,643682. Hull, J. and A. White, 1990, “ Pricing Interest-rate Derivative Securities”, Review of Financial Studies, 3, 573-592. Jones, C.M., Lamont, O., and Lumsdaine, R.L., 1998, “Macroeconomic News and Bond Market Volatility”, Journal of Financial Economics, 47, 315-337. Kadapakkam, P., and Kon, S., 1989, “The Value of Shelf Registration for New Debt Issues”, The Journal of Business, 62, 271-292. Kandel, S. and Stambaugh, R.F., 1996, “On the Predictability of Stock Returns: An Asset Allocation Perspective”, Journal of Finance, 51, 385-424. Kidwell, D; T. Koch, and D, Stock, 1984, "The Impact of State Income Taxes on Municipal Borrowing Costs", National Tax Journal, 37, 551-562. 31 Forbes, R. W. and P. A. Leonard , 1984, “The Effects of Statutory Portfolio Constraints on Tax-Exempt Interest Rates: Note”, Journal of Money Credit and Banking, 16, 92-99. Harris and Piowar, 2006, “Secondary Trading Costs in the Municipal Bond Market”, Journal of Finance, 61, 1361- 1397. Hong, G. and A. Warga, 2004, “ Municipal Marketability”, Journal of Fixed Income, 14, 86-95. Kadlec, G. B. and D.M. Patterson, 1999, “ A Transactions Data Analysis of Nonsynchronous Trading, Review of Financial Studies, 12,609-630. Lambert, L. and K. Pierog, 2009, “ U. S. Muni Bond Issuance Drops in 2009 First Half”, Thomson Reuters, June 30, 2009. Lewellen, J., and Shanken, J., 2002, “Learning, Asset-Pricing Tests, and Market Efficiency”, Journal of Finance, 57, 1113-1145. Lo, A. and A. C. MacKinlay ,1992, “ Maximizing Predictability in the Stock and Bond Markets”,Working Paper, Sloan School, MIT and Wharton School. Lo, A. and A. C. MacKinlay ,1997, The Econometrics of Financial Markets, Princeton University Press, Princeton, NJ. Lo, A. and A. C. MacKinlay ,1999, A Non-Random Walk Down Wall Street, Princeton University Press, Princeton, New Jersey. Marquering, W., and Verbeek, M., 2004, “The Economic Value of Predicting Stock Index Returns and Volatility”, Journal of Financial & Quantitative Analysis, 39, 407-429. Neiderhoffer,V. and Osborne, M.F. ,1966, “Market Making and Reversal on the Stock Exchange” Journal of American Statistical Association, 61, December, pages 897916). Papageorgiou, N. and F. S. Skinner, 2002, “Predicting the Direction of Interest Rate Movements”, Journal of Fixed Income, 11, 87-95. Pesaran, M.H. and A.G. Timmerman,2004, “ How Costly Is It to Ignore Breaks When Forecasting the Direction of a Time Series?” International Journal of Forecasting, 20, 411-425. 32 Schotman, P. C. and A. Zalewsaka, 2005, “Nonsynchronous Trading and Testing for Market Integration in Central European Emerging Markets”, Journal of Empirical Finance, 13,462-494. Securities Industry and Financial Markets Association (Outstanding U.S. Bond Market Debt, 2009). Vasicek, O. ,1977, An Equilibrium Characterization of the Term Structure”, Journal of Financial Economics, 5, 177-188. Xu, Y., 2004, “Small Levels of Predictability and Large Economic Gains”, Journal of Empiricial Finance , 11,65-77. 33 Figure 1 An analyst observes the close of the index at close t-1 and predicts the close for future periods. All the analyst knows is the closing index yield for t-1 and the predictions for future index closing yields. The prediction for closing index on day t is from an econometric model of index data. If the expected (predicted) index close for t is lower than t-1 index close, a downward trend is expected to occur for the window of issuance. So, the strategy is to delay issuance until t+1. If the predicted index close is higher (or the same) for t than the t-1 index, issue on day t, the actual day of issuance. Hopefully this issuance occurs before much of the predicted rise at close of day t (and the issuance window) occurs. Bloomberg reports issuance cost on day t (yhp,issue) where the actual time of issuance is sometime between open of t and close of t. Bloomberg does not report a closing yield for the individual bond on day of issuance, t, but, again, only the issuance yield (yhp,issue) which occurs sometime between the open and close of day t. Bond issued some time on day t at yhp,issue (the Bloomberg HP yield) Index Yields Bloomberg HP yields t-1 t t+1 close close t+2 close close Realized Closing Index on t-1 Predicted Closing Index on t Predicted Closing Index on t+1 yt-1 E(yt) E(yt+1) (nonexistent) yhp,issue yhp,t+1 yhp,t+2 (no closing day t yield supplied) 34