WIXLIB

The first task in implementing the MDE procedure is to specify the factors that determine MBS prices. To price MBS we need factors that capture the value of fixed cash flowcomponent of MBS and refinancing incentives. The particular factors we use here are theyield on 10-year Treasury notes and the spread between the 10-year yield and the 3-monthT-bill yield. There are good reasons to use these factors for capturing the salient features ofMBS. The MBS analyzed in this paper have 30 years to maturity; however, due to potential prepayments and scheduled principal repayments, their expected lives are much shorter.Thus, the 10-year yield should approximate the level of interest rates which is appropriatefor discounting the MBS’s cash flows. Further, the 10-year yield has a correlation of 0.98with the mortgage rate (see Table 1B and Figure 1). Since the spread between the mortgagerate and the MBS’s coupon determines the refinancing incentive, the 10-year yield shouldprove useful when valuing the option component.The second variable, the slope of the term structure (in this case, the spread between the10-year and 3-month rates) provides information on two factors: the market’s expectationsabout the future path of interest rates, and the variation in the discount rate over short andlong horizons. Steep term structure slopes imply lower discount rates for short-term cashflows and higher discount rates for long-term cash flows. Further, steep term structures mayimply increases in future mortgage rates, which should decrease the likelihood of mortgagerefinancing.2.2Multivariate Density Estimation IssuesThis subsection explains the details of the multivariate density estimation technique proposedin this chapter. To understand the issues involved, suppose that the error term in equation(1) is uniformly zero and that we have unlimited data on the past history of MBS prices.Now suppose that we are interested in determining the fair price for a MBS with a particularcoupon and prepayment history at a particular point in time when, for example, the 10-yearyield is 8% and the slope of the term structure is 1%. In this case all we have to do is lookback at the historical data and pick out the price of an MBS with similar characteristics at apoint in time historically when the 10-year yield was 8% and the slope of the term structurewas 1%. While this example illustrates the simplicity of underlying idea behind the MDEprocedure, it also highlights the sources of potential problems in estimation. First of all,for reasons discussed in the last subsection, it is unrealistic to assume away the error terms.Secondly, in practice we do not have unlimited historical data, and a particular economic5

scenario, such as an 8% 10-year yield and a 1% term structure slope, may not have beenplayed out in the past. The estimation technique therefore should be capable of optimallyextracting information from the available data.The MDE procedure characterizes the joint distribution of the variables of interest, inour case the joint distribution of MBS prices and interest rate factors. We implement MDEusing a kernel estimation procedure.2 In our application, the kernel estimator for MBS pricesas a function of interest rate factors simplifies to:PTt 1P̂mb,c (rl , rl rs ) Pmb,c,t KPTt 1 K rl rl,thrlrl rl,thrl KK [rl rs ] [rl,t rs,t ]hrl rs[rl rs ] [rl,t rs,t ]hrl rs ,(2)where T is the number of observations, K(·) is a suitable kernel function and h is the windowwidth or smoothing parameter. P̂mb,c (rl , rl rs ) is our model price for a MBS with couponc when the long rate is rl and the term structure slope is rl rs . Pmb,c,t is the market priceof the tth observation for the price of a MBS with coupon c. Note that the long rate at thetime of observation t are is rl,t and the term structure slope is rl,t rs,t .The econometrician has at his or her discretion the choice of K(·) and h. It is importantto point out, however, that these choices are quite different from those faced by researchersemploying parametric methods. Here, the researcher is not trying to choose functional formsor parameters that satisfy some goodness-of-fit criterion (such as minimizing squared errorsin regression methods), but is instead characterizing the joint distribution from which thefunctional form will be determined.One popular class of kernel functions is the symmetric beta density function, whichincludes the normal density, the Epanechnikov (1969) “optimal” kernel, and the commonlyused biweight kernel as special cases. Results in the kernel estimation literature suggest thatany reasonable kernel gives almost optimal results, though in small samples there may bedifferences (see Epanechnikov (1969)). In this paper, we employ an independent multivariatenormal kernel, though it should be pointed out that our results are relatively insensitive tothe choice of kernel within the symmetric beta class. The specific functional form for theK(·) that we use is:11 2K(z) (2π) 2 e 2 z ,2For examples of MDE methods for approximating functional forms in the empirical asset pricing literature, see Pagan and Hong (1991), Harvey (1991) and Ait-Sahalia (1996). An alternative approach toestimating nonlinear functionals in the derivatives market is described by Hutchinson, Lo and Poggio (1994).They employ methods associated with neural networks to estimate the nonlinear relation between optionprices and the underlying stock price.6

where z is the appropriate argument for this function.The other parameter, the window width, is chosen based on the dispersion of the observations. For the independent multivariate normal kernel, Scott (1992) suggests the windowwidth 1ĥi ki σ̂i T m 4 ,where σ̂i is the standard deviation of the ith variable (i.e., i may denote either variable rlor rl rs ), m is the dimension of the variables, which in our case is 2, and ki is a scalingconstant often chosen via a cross-validation procedure. In our application we need to chosetwo such scaling constants, one for the long rate rl and one for the term structure sloperl rs . Note that the window width is larger when the variance of the variable underconsideration is larger in order to compensate for the fact that observations are, on average,further apart. This window width (with ki 1) has the appealing property that, for certainjoint distributions of the variables, it minimizes the asymptotic mean integrated squarederror of the estimated density function. Unfortunately, our data are serially correlated andtherefore the necessary distributional properties are not satisfied.We employ a cross-validation procedure to find the ki that minimizes the estimation error.To implement cross-validation, the implied MDE price at each data point is estimated usingthe entire sample, except for the actual data point and its nearest neighbors.3 We identifythe ki ’s that minimize the mean-squared error between the observed price and the estimatedkernel price. Once the ki ’s are chosen based on cross-validation, the actual estimation of theMBS prices and analysis of pricing errors involves the entire sample.To gain further intuition into the estimation procedure, note that equation (2) takes aspecial form; the estimate of the MBS price can be interpreted as a weighted average ofobserved prices:P̂mb,c (rl , rl rs ) whereKwr (t) PT rl rl,thr t 1 K l Krl rl,thr l TXt 1wi (t)Pmb,c,t ,[rl rs ] [rl,t rs,t ]hr r K (3) sl [rl rs ] [rl,t rs,t ]hr r sl .Note that to determine the MBS price when the interest rate factors are (rl , rl rs )the kernel estimator assigns to each observation t a weight wr (t) that is proportional to the3Due to the serial dependence of the data, we performed the cross-validation omitting one year of data,i.e., six months in either direction of the particular data point in question.7

“distance” (measured via the kernel function) between the interest rate factors at the time ofobservation t (rl,t , rl,t rs,t ) and the current interest rate factors. The attractive idea behindMDE is that these weights are not estimated in an ad hoc manner, but instead depend onthe true underlying distribution (albeit estimated) of the relevant variables. Thus, if thecurrent state of the world, as measured by the state vector (rl , rl rs ), is not close to aparticular point in the sample, then this sample price is given little weight in estimatingthe current price. Note, however, that MDE can give weight (possibly inconsequential) toall observations, so that the price of the MBS with (rl , rl rs ) also takes into accountMBS prices at surrounding interest rates. This will help average out the different errorsin equation (1) from period to period. Although our application utilizes only two factors,MDE will average out effects of other factors if they are independent of the two interest ratefactors. Thus, for any given long rate rl and a given short rate rs , there is a mapping tothe MBS price Pmb (rl , rl rs ). These prices can then be used to evaluate how MBS pricesmove with fundamental interest rate factors.While the MDE procedure has the advantage that it does not require explicit functionalspecification of interest rate dynamics and prepayment models, it does have certain drawbacks. The most serious problem with MDE is that it is data intensive. Much data arerequired in order to estimate the appropriate weights which capture the joint density function of the variables. The quantity of data which is needed increases quickly in the numberof conditioning variables used in estimation. How well MDE does at estimating the relationbetween MBS prices and the interest-rate factors is then an open question, since the noisegenerated from the estimation error can be substantial.4Another problem with MDE is that the procedure requires covariance stationarity of thevariables of interest. For example, when we use only two interest rate factors, the MDEprocedure does not account for differences in prices MBS when the underlying pools havedifferent prepayment histories. For this reason the MBS procedure is most suitable forpricing TBA securities which are most commonly used for new originations rather than forseasoned MBS. Accounting for seasoning of a mortgage or a mortgage pool’s burnout willrequire additional factors that are beyond the scope of this chapter.A few comments are in order, however, to provide some guidance on how these factorscould be accounted for when one is interested in pricing seasoned MBS. First, one could4Boudoukh, Richardson, Stanton and Whitelaw (1997) perform simulation exercises in an economy governed by two factors and some measurement error in reported prices. Within this (albeit simple) environment,the MDE methodology performs quite well.8

potentially take account of a mortgage pool’s seasoning by nonlinearly filtering out any timedependence. Estimation error aside, this filtering would be effective as long as the seasoningis independent of the other state variables. Second, in order to incorporate path dependencedue to a pool’s burnout, the only viable way would be to employ a state variable whichcaptures this dependence. For example, Boudoukh, Richardson, Stanton and Whitelaw(1997) and Richard and Roll (1989) describe several variables that might be linked closelywith burnout. Because the strength of the MDE procedure estimation of nonlinear relations,all that is required is that these variables span the appropriate state space.3Data Description3.1Data SourcesMortgage-backed security prices were obtained from Bloomberg Financial Markets coveringthe period January 1987 to May 1994. Specifically, we collected weekly data on 30-yearfixed-rate Government National Mortgage Association (GNMA) MBS, with coupons rangingfrom 7.5% to 10.5%.5 The prices represent dealer-quoted bid prices on GNMAs of differentcoupons traded for delivery on a to be announced (TBA) basis.The TBA market is most commonly employed by mortgage originators who have a givenset of mortgages that have not yet been pooled. However, trades can also involve existingpools on an unspecified basis. Rules for the delivery and settlement of TBAs are set by thePublic Securities Association (PSA) (see, for example, Bartlett (1989) for more details). Forexample, an investor might purchase 1 million worth of 8% GNMAs for forward deliverynext month. The dealer is then required to deliver 8% GNMA pools within 2.5% of thecontracted amount (i.e., between 975,000 and 1,025,000), with specific pool informationto be provided on a TBA basis (just prior to settlement). This means that, at the time of theagreed-upon-transaction, the characteristics of the mortgage pool to be delivered (e.g., theage of the pool and its prepayment history) are at the discretion of the dealer. Nevertheless,for a majority of the TBA’s, the delivered pools represent newly issued pools.With respect to the interest rate series, weekly data for the 1987-1994 period werecollected on the average rate for 30-year mortgages (collected from Bloomberg Financial5Careful filters were applied to the data to remove data reporting errors using prices reported in the WallStreet Journal. Furthermore, data are either not available or sparse for some of the GNMA coupons duringthe period. For example, in the 1980’s, 6% coupon bonds represent mortgages originated in the 1970’s, andnot the more recent issues which are the focus of this paper. Thus, data on these MBS were not used.9

Markets),6 and the yields on the 3-month Treasury bill and 10-year Treasury note (providedby the Board of Governors of the Federal Reserve).3.2Data CharacteristicsBefore describing the pricing results and error analysis for MBS using the MDE approach,we briefly describe the environment for interest rates and mortgage rates during the sampleperiod, 1987-1994.Characteristics of Mortgages (1987-1994)Since the mortgage rate represents the available rate at which homeowners can refinance,it plays an especially important role with respect to the prepayment incentive. Figure 1graphs the mortgage rate for 1987 through 1994. From 1987 to 1991, the mortgage ratevaried from 9% to 11%. In contrast, from 1991 to 1994, the mortgage rate generally declinedfrom 9.5% to 7%.7For pricing GNMA TBAs, it is most relevant to understand the characteristics of theuniverse of pools at a particular point in time. That is, the fact that a number of poolshave prepaid considerably may be irrelevant if newly originated pools have entered into theMBS market since the MBS from new originations are the one typically delivered in TBAcontracts. To get a better idea of the time series behavior of the GNMA TBAs during thisperiod, Figure 2 graphs an artificially constructed index of all the originations of 7.5% to10.5% GNMA pools from January 1983 to May 1994.8There is a wide range of origination behavior across the coupons. As mortgage ratesmoved within a 9% to 11% band between 1987 to 1991, Figure 2 shows that GNMA 9s, 9.5s,10s and 10.5s were all newly originated during this period. Consistent with the decline inmortgage rates in the post 1991 period, GNMA 7.5s, 8s and 8.5s originated while the GNMA9s–10.5s became seasoned issues. Thus, in terms of the seasoning of the pools most likelyto be delivered in the TBA market, there are clearly cross-sectional differences between thecoupons.6Bloomberg’s source for this rate is “Freddie Mac’s Primary Mortgage Market Survey”, which reportsthe average rate on 80% of newly originated 30-year, first mortgages on a weekly basis.7Note that the MBS coupon rate is typically 50 basis points less that the interest rate on the underlyingmortgage. The 50 basis point is retained to cover the servicing fee and reinsurance cost.8The dollar amount outstanding for each coupon is normalized to 100 in January 1987. Actual dollaramounts outstanding in that month were 10,172, 27,096, 10,277, 63,392, 28,503, 15,694, and 5,749(in millions) for the 7.5% – 10.5% coupons, respectively.10

Figure 2 shows that there are several reasons for choosing the TBA market during thepost 1986 time period to investigate MBS pricing using the MDE methodology. First, during1985 and 1986, interest rates dramatically declined, leading to mortgage originations for awide variety of coupon rates. Thus, the GNMA TBAs in 1987-1994 correspond to mortgagepools with little prepayment history (i.e., no burnout) and long maturities. In contrast,prior to this period, the 7.5% to 10.5% GNMAs were backed by mortgages originated in the1970’s and thus represented a different security (in both maturity and prepayment levels).Second, MDE pricing requires joint stationarity between MBS prices and the interest ratevariables. This poses a potential problem in estimating the statistical properties of any fixedmaturity security, since the maturity changes over time. Recall that the TBA market refersto unspecified mortgage pools available in the marketplace. Thus, to the extent that thereare originations of mortgages in the GNMA coupon range, the maturity of the GNMA TBAis less apt to change from week to week. Figure 2 shows that this is the case for the highercoupon GNMAs pre 1991, and for the low coupon GNMAs post 1991. Of course, when nooriginations occur in the coupon range (e.g., the GNMA 10s in the latter part of the sample),then the maturity of the available pool will decline. In this case, the researcher may needto add variables to capture the maturity effect and possibly any prepayment effects. In ouranalysis, we choose to limit the dimensionality of the multivariate system, and instead focuson the relation between MBS prices and the two interest rate factors.Characteristics of MBS Prices and Interest Rates(1987-1994)Table 1 provides ranges, standard deviations and cross-correlations of GNMA prices(Table 1A), and mortgage and interest rates (Table 1B) during the 1987-1994 period. Absentprepayments, MBS are fixed-rate annuities, and the dollar volatility of an annuity increaseswith the coupon. In co

discount bond prices. The complexity in pricing of MBS is due to the fact that statutorily mortgage holders have the option to prepay their existing mortgages; hence, MBS investors are implicitly writing a call option on a corresponding xed-rate bond. The timing and magnitude of cash flows from MBS are therefore uncertain. While mortgage .