Prepayment Risk- and Option-Adjusted Valuation of MBS

opportunities for arbitrai^e.

Alexander Levin and Andrew Davidson

ALEXANDER LEVIN is head

ot Valuation i^cvclopment at Andrew Davidson &: Co., Inc., in New York City.

[email protected]
A N D R E W D A V I D S O N IS the

president and founder of Andrew Davidson & Co., Inc.

[email protected]

ption-adjusted spread (OAS), while a much better measure than yield or static spread, still tails short in explaining the dynamics of mortgage pricing. The standard OAS typically varies across instruments {pass-throughs, collateralized mortgage obligations, interest-only securities, principal-only securities), coupons, prepayment option moneyness, and pool seasoning stages. Premium and discount MBS are often priced at wider OAS than the current-coupon issues. Premium MBS and lOs stripped oft premium pools are considered hazardous, and their higher OAS reflect concerns of understated refinancing. Naturally, the respective POs look rich. In the discount sector, higher OAS reflects the risk associated with possible overstatement of the housing turnover rate. Clearly, these market phenomena defeat the very purpose ofa constant OAS approach. Rich-cheap judgments become inconclusive, and rate shock analysis can produce inaccurate hedge ratios.' Like Cheyette [1996] and Cohler, Feldman, and Lancaster [1997], we attribute the OAS and its variability to the prepayment risk premium, i.e., possible nondiversifiable deviations of actual future prepayments fix)m a best guess prepay model's forecast. It is the market's fear of systematic bias in prepayment forecasts that leads to a risk premium. If prepayments were perfectly predictable, then an exact, even inefficient, option exercise model should produce a zero OAS to an appropriate option-trf e benchmark, just as options and

SUMMKR 2005

THE JOURNAL OF PORTFOHO MANAGEMENT

73

embedded option instruments with known algorithms of exercise such as swaptions, callable agency debt, and corporate bonds are all priced flat to their respective optionfree rate curves. While OAS varies widely among instruments, our new spread measure, called prepayment risk- and optionadjusted spread (prOAS, pronounced pro-A-S), accounts for both option and prepayment risk. We posit that, on a prOAS basis, all liquid agency MBS should he priced flat to agency debentures, eliminating the variability found in traditional OAS measures. Our method has its roots in the capital asset pricing model (CAPM) and its extension, arbitrage pricing theory (APT), where return compensation for risk and the concept of equivalent risk-neutrality play key roles. For unstructured pass-throughs, we propose a prOAS valuation model that is armed with the power of backward induction and allows for endogenously finding risk measures and prices reflecting embedded prepay uncertainty^in the form of return spread compensation computed for each investment period and level of interest rates. All the required prices and risk spreads can be found concurrently in the course of valuation performed backward on a probability tree or a fmite-difference grid. A price obtained in this manner will reflect the differences in prepayment uncertainty without the need to vary OAS across instruments. Values of IOs, POs, and mortgage servicing rights (MSR) can be objectively derived without knowing the OAS level for each instrument; if necessary, traditional OAS can later be calculated from the resulting prices. We prove that this process of explicit endogenous prepay risk accounting is mathematically equivalent to risk-neutral prepayment modeling. Such a model retains the structure and the features of a physical prepay model, but operates with risk factors stressed to their undesired directions (where value deteriorates). A risk-neutral prepay model easily solves the issue of C M O valuation under price of prepay risk implicitly, without reliance on a non-feasible backward valuation. Tbat is, we could refine the handling of CMOs hy replacing the time-consuming prepayment stress tests described in Cohler, Feldman, and Lancaster [1997] witb risk-neutral forecasts. We find that two prepayment risk factors are essential: the risk of refmancing understatement, and tbe risk of turnover overstatement. Having calibrated prices of these two prepayment risks to a set of widely traded MBS, we can then produce prices for all other instruments exposed to the same risks. Without two independent risk 74
PREi'AYMENT RiSK- AND Ol'TION-AnjUSTEH VALUATION OF MBS

factors, it would be impossible to explain why botb discounts and premiums of nearly all MBS collateral types are traded at higher OAS tban non-MBS instruments. This paradox bas apparently puzzled some authors; see Kupiec and Kah [1999] and Gabaix, Krishnamurtby and Vigneron [2004]. A single-dimensional risk analysis would allow for hedging prepayment risk by combining premium MBS and discount MBS, a strategy any experienced trader knows would fail. PRICING FDE FOR OPTION-ADJUSTED VALUATION We start with a hypothetical dynamic asset (mortgage) whose market price P(r, .v) depends on time f and one market factor .\'. We treat x{t) as a randotn process with a (generally variable) drift rate /i and a volatility C , disturbed by T a standard Brownian motion z{t):

dx ^

+ adz

(1)

We assume further that tbe asset pays tbe continuous coupon rate c{t, x), and its balance B is amortized at tbe A(f, .v) rate, so that (dB/dt) = ~XB. Then, one can prove that tbe price function P{t, x) should solve the partial diflx'rential equation (PDE):

r -+ OAS = - + (c + A) - A + dP

2Pdx-

(2)

A derivation of this PDE can be found in Levin [1998], but it goes back at least to Fabozzi and Fong [1994[. If our amortizing asset is an interest-only strip (IO), the pricing equation is modified by excluding the X/P term in the time return expression. A notable feature of the PDE (2) is tbe absence of tbe balance variable, JB. Tbe entire effect of possibly random prepayments is represented by the amortization rate function, X{t. x). Although the total cash flow observed for each accrtial period does depend on tbe beginning-period balance, construction of a backward induction scheme will require the knowledge of A(r, x), not the balance. Pricing PDE (2) can be solved on a probability tree or a fmite-difference grid tbat has as many dimensions as the total number of factors or state variables r, c, and A,
SUMMER 2005

or by Monte Carlo simulation. If the coupon rate is fixed, and the amorrization rate A depends only on current time {loan age) and the immediate market factor x, the entire valuation problem can he solved backward on a twodimensional {x, l) lattice (the lattice will require more dimension if the market factor .v is a vector). To implement this method, we would start the valuation process trom maturity 7'when wt* surely know that the price is par, P{T. x) 1 (zero for an IO), whatever the value of factor x. Working backward, we deduce prices at age f - 1 from prices found at age t. In doing so, we replace derivatives in PDE (2) by finite-difference approximations, or weigh branches of the lattice by explicitly computed probabilities. Even for a simple fixed-rate mortgage pass-through, though, total amortization speed A usually cannot be modeled as a function of time and the immediate market. Prepayment burnout is a strong source of path-dependence because tuture refinancing activity is afiected by the past incentives. One can think ofa mortgage pool as a heterogeneous population of participants with different refinancing propensities. Some mortgagors have higher rates or better credit or larger loans, or perhaps they face lower regional transaction costs. Once they leave the pool, future prepayment activity gradually declines. Hence, A depends on historical market rates, making the valuation problem path-dependent. Instead of considering pricing PDE (2) for the entire collateral. Levin [2001, 2004] proposes decomposing it into two components, active and passive, differing in refinanceability. Under the active-passive decomposition (APD) model, mortgage path-dependent collateral can be deemed a simple portfolio of two path-independent instruments if; Active and passive components prepay differently, but follow immediate market rates and loan age. Any migration between components is prohibited. Alternatives and variations of this modeling structure include decomposition into several groups, as well as stratifying the entire pool explicitly by available loan characteristics (see Davidson [1987], Davidson, Herskovitz,aiid Van Drimen[1988|, Hayre |1994, 2()0{)], and Kalotay, Yang, and Fabozzi [2004]). With a decomposed collateral, it is possible to value mortgage-backed securities using backward induction and, as we show further, explicitly account for prepayment risk.

VALUATION WITH PREPAYMENT RISK: BASIC CONCEPTS Mortgage practitioners use the term prcpayntctu risk loosely. Most often they simply mean prepayment variability, but this is not what we attempt to capture. Indeed, a large portion of prepayment uncertainty is associated with interest rates and is thus explained by the prepayment models that are inherent in modern option-adjusted spread analytical systetns. If prepayments were perfectly explained by a model, there would not be any prepayment risk premium. An option model coupled with an exact prepayment formula should be ahle to deliver the right price for an agencybacked (default-protected) MBS operating with OAS = 0. MBS would be valued flat to a known benchmark curvesimilarly to swaptions or callable notes, except with a more complex exercise rule. Savvy market participants realize that a model can tell only part of the prepayment story. Because a model cannot predict prepayments exactly, we have unexplained deviations of prepayment speeds above or below the model's forecast, often called prepayment surprises or prepayment errors. Not all prepayment surprise should require market compensation. Random oscillations of actual prepayments around the model are diversifiable over time. Prepayment errors typically seen in small pools are diversifiable in large pools. We associate the notion of prepayment risk with mmdiuersifiabU' niia-nainty. common for the MBS market, systematic in trend and unexplained by an otherwise bestguess prepay model. The varying level of OAS on different instruments represents compensation for this risk. Al! the terms in pricing Equation (2) represent different sources of return, but none of them explicitly quantifies prepayment risk. The entire compensation for bearing this risk is hidden in the OAS term. How would we price prepayment risk? Suppose that the prepayment rate A (r, x, i^), depends on one uncertain variable or uncertain parameter, ^, independent of the interest rate market. For the first conceptual illustration, we assume ^(r) is a Wiener process with zero drift and volatility of <7c, i.e., d^= G^dz^, and known initial value, <^(()). According to the CAPM/APT, the risky return should be proportional to the price volatility due to the risky factor ^. A comnion multiplier, Tie, called price of risk, should apply to every asset exposed to the same risk factor ^. Using this notation, we therefore state that, for every

SUMMER 2(III5

THE JOURNAL OF PORTFOHO MANAGEMENT

75

investment period, the expected return, r + OAS on the left-hand side of PDE (2), should be adjusted tor risk as: Single-Period Expected Return

p (r + prOAS)P - at

price of iTik

(4)

prOAS+

idP
prepay nsk ipread

(3)

Second, we take the first derivative of both sides of Equation (4) with respect to t^: dP 1

where the prepay risk- and option-adjusted spread (prOAS) is a "risk-free" OAS; in the absence of any other risk factors, it should be zero for a properly selected pricing benchmark. For example, we may assume that all agency MBS should be valued flat to the same agency yield curve, on this prepay risk-adjusted basis (prOAS = 0). For nonagency MBS, this should certainly account for an additional risk associated with imperfect credit, and thus becomes equal to the pure credit spread that can be derived from the S&P or Moody's rankings for a nonagency pass-through or a particular C M O tranche. The risky spread term, unlike the traditional OAS. is not constantit varies with interest rates and loan age. It is also directionalit can be both positive and negative depending on the sign ot price exposLire to the factor <^. Since the market provides a return premium tor bearing the risk, it must make hedge instruments rich so that a properly constructed i^-neutral portfolio of an asset and its hedge will earn nothing but risk-free OAS, i.e., prOAS. This is a traditional arbitrage argument that prevents construction of a risk-free portfolio earning any excess above the risk-free return (see Hull |2000]). VALUATION METHOD 1: ENDOGENOUS ASSESSMENT OF RISK Factor volatility (Te and price of risk constant K^ are defined outside the pricing model; they are common tor all instruments exposed to prepay tactor ^. Let us showhow price P(/, x\ ^) and its partial derivative taken with respect to the risk factor, dP/d^, denoted trom now on as Pc, can be estimated internally and simultaneously in the course of backward valuation. Suppose we have a path-independent instrument in hand that is subject to pricing Equation (2), and the amortization rate X depends on risk factor ^. Let us first multiply both sides of pricing Equation (2) by price P and replace the traditional expected return. r + OAS, with the risk-adjusted version from (3): 76
PREPAYMENT RrsK- AND O[>TioN-AiiiusTEi.i VALUATION OF MBS

(r -f prOAS)P. ^

(1 - P)term (o oniit or approximate

(5)

where for now we disregard the last term. We have here a system of two linear partial differential Equations (4) and (5) with two unknown functions, P((, A) and PM, x). It can be simultaneously solved " backward starting from the terminal (maturity 7) conditions: P{T, x) - 1, PA.T, x) - 0. This backward induction is carried out on a usual {t, A")-grid or probability tree, for one value of (^ = ^(0), and separately for the active and passive part. Omission or approxinaation of the prepayment convexity term, proportional to P==, avoids expanding the pricing grid to the (^-dimension. If we want to account for the convexity term, we can do so by adding \Pif<yl to the PDE, and then differentiate both parts twfce while neglecting (or approximating) the third derivative. We would end up with a system of three PDEs, to be solved on the same ((, A:)-grid for three unknown functions, P[i, .v), P-(f, x). and P..(^ A")see Levin [2004] for details. VALUATION METHOD 2: EQUIVALENT RISK-NEUTRAL PREPAY MODEL Pricing Equation (4) allows for an important financial interpretation. Suppose we still work w^ith the traditional PDE (2), ignoring the price of risk but instead letting the risk tiictor B, drift with a negative Kg. (7= rate per year:

dt= a.dz. - Tt.G.dt

(6)

It is easy to see that such a drift contributes a systematic return that is mathematically identical to the above marked price of risk. Indeed, the full-time derivative term, dP/dt on the right-hand side ot Equation (2), will now be composed of a dP/dt term measured due to a simple pasSUMMER 20U5

sage of time {i.e., with unchanged ^ minus a TT^ C c Pc term T that comesfix)mthe Ito lemma applied to price P as a flinctioii of random variable ^ defined by Equation (6). This leads directly to the pricing Equation (4) with risk. We essentially arrive at a powerful risk-neutrality concept for prepayment risk: IVicing with risk consideration can be replaced with pricing without it, but with the prepay risk factor ^ drifting in the undesired direction. The rate ot this drift is proportional to volatility <Tc, and the coefficient of this proportionality is the price of risk, Kp. This IS the same concept we use in constructing term structure models and pricing financial derivatives wbilc taking forward rates and prices into consideration. A prepayment model with the factor ^ set to drift at the risk-adjusted rate can be logically called a risk-ncutrat prepayment model, like all other financial models that exploit this concept. It is therefore meaningless to wonder how well such a model fits actual prepaymentsit is simply not meant to do this job. The risk-neutral drift for factor ^ can also be partly caused by a model's systematic bias. For example, the market may expect tbe refinancing process to become more efficient going forward than it has been. Mathematically this bias cannot be distinguished from the price of risk; it too results in a drift cbange for ^. Using tbe riskneutral prepay model and the prOAS should lead to the same valuation results as using an empirical prepayment model witb the traditional OAS. Mortgage market participants are accustomed to seeing ditferences in broker and analyst forecasts of prepayments and reports of OAS numbers for the same instruments and market conditions. The transition 6x)m an objective model, which often Lises historical prepayments, to a risk-neutral one that targets known prices for known instruments may explain the disparities. Hence, risk-neutral prepay speeds should differ less across firms. Since Black and Scholes, risk-neutral modeling has been known for its abihty to reduce bias from systematic model-induced errors. Although we have shown that a risk-neutral prepay model is theoretically equivalent to the explicit risk assessment performed for each investment period, from a financial engineer's point of view there is a fundamental difference between the two techniques. The exphcit risk assessment method computes the partial derivative Pdirectly tor each investment period and rate level. This task is feasible if the entire valuation is performed backward, and the MBS price and its derivatives with respect to risky factor (^ can be found tor every node of the pricing grid. For example, unstructured MBS can be priced this
2(105

way using the burnout-curing active-passive decomposition idea. CMOs, though, are heavily path-dependent beyond burnout, and are therefore not subject to backward valuation. For CMOs, the risk-neutral prepay model would be an ideal method as it does not require computing Pp directly. Letting the risk factor ^ drift in the undesired direction naturally explores the price's dependence on (^, without measuring P^ explicitly. We have so far assumeci for simplicity that the risk factor follows a simple Wiener process. In general, the behavior of prepayment factors is more complex. For example, prepayments cannot grow more uncertain over time without bounds, suggesting that the dynamics of risk factors are mean-reverting. Yet purely diffusive behavior cannot capture uncertainty in tbe starting values, i.e., risk present at time zero. For example, the price ot risk constant model considers risky prepayment parameters where values are constant but drawn from some distribution (see Cohler, Feldman, and Lancaster [1997|). In general, we may extend the stochastic model for factor ^(r) to:

(7) VarBiO) - a; where d- is the mean reversion parameter, and ^ is the long-term equilibrium. The starting value is now considered uncertain and drawn from N]^^^, (7^J. This meanreverting pattern will add more terms in pricing PDEs. Also, tbe very last {time zero) step in backward valuation will become a special oneto account for the time zero risk and convexity. Transforming model {7) into a risk-neutral form, we again add a -K^CT,: drift rate, and shift the initial condition by -;rc(T^: G.dz {7-RN)

These are the generalized dynamics of the risk factor ^{t) in a risk-neutral prepayment model. We see that risk-neutrality lowers both the starting value and {in the presence of mean reversion) the long-term equilibrium for ^{/).

THE JOURNAL OF POUTFOLII) MANAGEMENT

77

A PROAS P R I C I N G MODEL WITH REFINANCING AND TURNOVER RISK

We have noted that both premium and discount MBS are often traded at somewhat elevated option-adjusted spreads. IOs stripped fixjm premium collateral are priced progressively cheaper, on an OAS basis, than IOs taken fi^om discount or current-coupon collateral. We conjecture that two major prepayment sources, the refinancing process (driving the premiums) and the turnover process (vital for understanding the discounts), are perceived as risky by the mortgage market. To put this into simple practical terms, there are two distinct market fearsrefinancing understatement, and turnover overstatement. Hence, the model of risk should be at least two-dimensional, which is a rather simple extension ot what we have considered so far. Allowing the refinancing and the turnover processes to be randomly scaled, we can model the total prepayment speed as: SMM - pRefiSMM + iTurnoverSMM (K)

where the prepay multipliers, p and T, are considered uncertain but centered on 1, and mutually independent. Thus, instead of one hypothetical prepay risk factor i^, we bave now two, p and T. If we use the active-passive (or other) collateral decomposition, we apply the additive rule in Equation (8) to each constituent piece. Every risk premium and convexity cost found in pricing equations now becomes a simple sum of two terms associated with the refinancing risk and the turnover risk. We assume known volatihties, (7 and (T^ of two Wiener processes, p{l) and r(f), as well as two prices of risk, 71 and 71^ The conceptual risk-adjusted return formula in Equation (3) now becomes Single-Period Expected Return = r -\- prOAS P, ^p
prepay risk spread

Extension ot the single risk factor is rather simple. In the fundamental pricing PDE (2), we now replace the expected return, r + OAS, with Equation (9). Thus modified risk-adjusted PDE will apparently include unknown derivatives Pj.and P . The next step is to differentiate this equation with respect to each risk factor, p and r, thereby adding two more equations and closing mathematical construct. The total number of pricing PDEs to solve will be either three with prepay convexity cost disregarded (for P, P , and P ), or six with prepay convexity cost included (for P, P , P^ the second derivatives, P , P and the mixed derivative P^^)-' For this two-factor prepay risk setting, a risk-neutral prepay model remains an attractive alternative method of computing prOAS. We accelerate refinancing (set p drifting above 1 at the rate ofu <JX and retard turnover (set T drifting below t at the rate of JC^CTj}. For a framework that combines single (time-zero) jump and meanreverting diffusion, these risk-neutral drifts tor the refmancing and the turnover multiples are shown in Exhibit 1. We can split the risk between jump and diffusion, and alter the rate of reversion. The dynamics shown in Exhibit 1 associate about half ot refinancing risk with the refinancing uncertainty present at time zero; the other half appears gradually. We are generally more certain about the starting turnover rate, and it may take a while before future macroeconomic conditions will alter it. so a smaller portion of the total drift is present at time zero for the turnover risk component. DETERMINING PRICES OE RISK: CALIBRATION TO TBAs Our discussion of practical steps m using the prOAS model includes the problems it may address and the valuation results it generates. Assuming we know volatility and mean-reversion parameters for the risk factors p{t) and r(f), we need to find market levels for (calibrate) the price of risk constants, K and 71^. We can tune these constants to match market prices (or the OAS), for a range of actively traded securities. To find these parameters, we first need to define the appropriate target for prOAS. The target prOAS should reflect the spread on a similar securit\' with no prepayment uncertainty. While assuming a zero prOAS to the swap curve may be appropriate, we fmd that agency debentures are a better benchmark, as they have the same credit guarantee as agency MBS and entail no prepayment uncertainty.'
SUMMER 21H)5

P.
f p

where P and P stand tor partial derivatives. Note the negative sign for the refinancing risk. Since premium tixed-rate MBS typically have P < 0, we can reward them by either assuming a negative price of risk constant, 71 , or using the negative sign in the spread formula. Discount fixed-rate MBS have P^ > 0, so the positive sign in (9) produces positive return compensation for bearing turnover risk. 78
PKEI'AYMENT RISK- AND OPTION-ADJUSTEP VALUATION OF MBS

EXHIBIT 1 Risk-Neutral Prepay Multiples

prepayment multiple 1.5

refinancing, p (t)

time-zero jumps recognizing uncertainty in starting vaiues physicai equilibrium

turnover, T (f )
0.7

6 7 years forward, t

10

11

12

An example of the calibration results is shown in Exhibit 2. We use eight TBA instruments priced on August 29, 2003, with net coupons ranging from 4.5% to 8.0%. On that day, the mortgage current coupon rate was 5.67%, so there were both premiums and discounts in our sample. First, we measure OAS numbers using the traditional valuation method, without any risk adjustment (black bars). We then use the prOAS pricing method and select the price of risk constants 71 and 71 so as to minimize the prOAS levels (grey bars). The calibration works fairly well across the range of TBAs, with a 3-4 basis point mean-squared accuracy in reaching the debenture (zero) prOAS target. The lines drawn in Exhibit 2 show the principal components of OAS, i.e., OAS compensation due to refinancing risk and turnover risk. The direction of both lines is apparent, but some important points should be made. For one, the turnover line almost never leaves positive territory. Discounts would certainly lose value with slow turnover, but why will premiums suffer? The very steep yield curve is primarily responsible for this effect; slowing turnover pushes cash flows to longer maturities
SUMMEU 2005

with higher discount rates. It also slightly inflates the time value of the prepayment option. The best mix of principal components is found assuming that prOAS is linear in prices of risk. Hence, our actually achieved prOAS levels are suboptimal (white bars). This calibration exercise clearly shows the value of the two-risk factor model for several reasons. In a singlerisk factor model, prepayment risk is associated with either global acceleration or deceleration, regardless of the source. First, in seeking a single price of risk, it would be impossible to move the premium OAS levels toward the zero level without moving the discount OAS levels away from zero, and vice versa. Second, for any market condition, a single-risk factor model would allow one MBS coupon {perhaps, interpolated) to be prepay-neutral, i.e., not exposed to the overall prepay scale. Hence, the theory requires such an MBS be traded at a zero, not a positive, OAS, which is not the case for the market shown in Exhibit 2. Third, as we mentioned in the introduction, a single-factor risk model leads to an unworkable singlefactor hedge strategy. Will the parameters of the prOAS model be stable
THE JOURNAL OF PORTFOLIO MANAGEMENT 79

EXHIBIT 2 Prices of Risk Calibrated to FNCL TBAs as of August 29, 2003

Agency OAS/prOAS, bp 60 usual OAS PCA-optimized prOAS (RMSE = 2.9 bp) actually achieved prOAS (RMSE = 3.6 bp) refinancing principal component turnover principal component

FNCL4.5

FNCLS.O

FNCL5.5

FNCL6.0

FNCL6.5

FNCL7.0

FNCL7.5

FNCLS.O

over time, or do we need to calibrate them on a daily basis? Wbile one goal of physical models is to stay steady, the concept of risk-neutrality is linked to changing market prices for benchmark instruments, which reflect the dynamics of market preferences for risk. If the market prices for TBAs exhibit OAS tightening or widening over time, they are sending us a message ot changing perceptions of prepayment risk. This conjecture is borne out when we examine the trends in results of the calibration of tbe prices of risk constants, 71 and Jlj^ at different dates, as shown in Exhibit 3. These parameters are not constant and even show an exaggerated reaction to interest rate dynamics. When rates dropped to tbeir 4(.)-year record lows (May-June 2003), refinancing fears reached panic stage. High-premiums (e.g., FNCL7.5 and FNCLS.O) did not appreciate, which meant that their agency OAS levels increased to 100 basis points and above to absorb much of the rate plunge. During that period, the calibrated price of refinancing risk surged. The calibration revealed no concern about housing turnover, as the discount sector had evaporated. When rates moved back up through summer 2003, the refinancing wave started to cool off; large volumes of 80
PREPAYMENT RISK- AND OpnoN-AoiusTEn VALUATION OF MBS

freshly originated FNCL4.5 and FNCL5.0 became discounts. This was when the turnover concerns became apparent. The rest of the time we witnessed a general stabihzation in risk prices. In our opinion, it is dynamics of interest rates, not their levels, that induce exaggerated prices of risks. Both irrational pricing and sharp changes in the mortgage market composition may explain this phenomenon (see also Gabaix, Krishnamurthy, and Vigneron [2004]). Comparing the heights of the bars in Exhibit 3, one might conclude that the MBS market is systematically dominated by refinancing fears, not turnover fears. This is likely a misperception. As Exhibit 2 shows, tbe principal components of OAS are of comparable size even when K is significantly greater tban K^ When rates rise, prices for discount MBS can suffer substantial drops that are limited by and therefore depend on the turnover speed. VALUATION OF MBS STRIPS WITH
PROAS

In the traditional OAS valuation, either price or OAS should be given as input. Under prOAS, the role of OAS is performed by a better-defined measure. The goal of prOAS pricing is to eliminate differences in OAS
SUMMER 2005

EXHIBIT 3 Historical Prices of Refinancing and Turnover Risk

Price of Risk 4.5 Rats, -6

refinancing scare

turnover concerns

stabilization
5.S ^m Refl Risk MTGEFNCL 5.4

t- . I Turnover Risk

5.6

4.6

0.5

4.4

4.2 Apr-03 -0.5 May-03 Jun-03 Jul-03 Aug-03 Sep-03 Oct-03 Nov-03 Dec-03 Jan-04 Feb-04 Mar-04

among instruments that are exposed to prepayment risk. As we asserted earlier, the prOAS measure should value agency MBS flat to agency debentures. Therefore, once the risk factors are given their stochastic specifications (jump sizes, diffusive volatility, and mean reversion) and the prices of risk constants are determined, we can value any agency MBS or its IO and PO strips much like swaptions, i.e., hy looking at the benchmark rate and volatility structure, but without any knowledge of the traditional OAS. Exhibit 4 shows valuation results for agency trust IOs using prices of risk constants, a high T , and a near-zero T 7Cj., obtained from the cahbration to Fannie Mae TBAs on May 30, 2003. Application of the prOAS method first leads to values that are then converted into conventional OAS measures.^ The prOAS model explains IO cheapness (and therefore PO richness) naturally, and correctly predicts an OAS level of (and above) 1,000 basis points. As POs stripped off premium pools should be looked at as hedges against refinancing risk, they have to be traded rich according to arbitrage pricing theory. Our prOAS model successfully confirms this in that virtually all OAS for trust POs are deep in negative territory (not shown in Exhibit 4). Results in Exhibit 4 also provide
SUMMER 2005

some degree of confidence for managers of mortgage servicing rights (MSR) portfolios (not actively traded or frequently quoted)they can use the prOAS measure to better assess the risk of their portfolios. Exhibit 3 sbows in a historical risk chart that on May 30, 2003, the entire risk perception was evolving out of a refinancing scare. Exhibit 4 shows that both the TBA market and the trust IO market agree with one another in incorporating this risk into pricing. In the summer months of 2003, rates rose sharply, pushing lower-coupon MBS (4.5s and 5.0s) into discount territory. Exhibit 2 confirms that by the end of that summer the refinancing fear had dissipated, making room for turnover concerns (slower-than-modeled turnover results in a loss for a discount MBS holder). It is not surprising that the price for turnover risk, virtually non-existent in May 2003, grew considerably (Exhibit 3). What if we apply prices of risk calibrated to the August 29, 2003, TBA market to value trust IOs? Exhibit 5 shows two stages in application of the prOAS model, valuation with refinancing risk only, and with total risk. Comparing the market prices and related OAS (light grey lines) with the valuation results under the prOAS model with refinancing risk only {darker grey hnes), we
THEJOURNAI OF PORTFOLIO MANAC;F.MI;NT 81

EXHIBIT

4
OAS, bps 4000 6.0 6.5 7.0 7.5 8.0

Trust IO Valuation as of May 30, 2003

Value, % points
20 5.5

18 16
14 ^ ^ market quotes

s
i
dr

3500

h \ \
\
V

^^^derived with prOAS

3000

12 10

2500

Values

^ V ^

ft
C
1

l\
_ .-/I

2000

1500

1 1
1 ^'

OAS

1000

^*l 1

500

izzzrizzxrzzzrizzzzrzzriziziiizzrzziziizriizizzi

see that the single-risk factor prOAS model just slightly overstates values compared to the actual market. It shows a directionally correct OAS tendency (tighter spread for discount IOs, wider for premiums) and magnitude. Disaster strikes when we add the true turnover risk calibrated to the TBAs (black lines). Since IOs can be used as hedges against turnover risk, theory says they should be penalized, not rewarded. An almost constant 250-basis point OAS reduction is seen as the result. On that day the actual IO market did not seem to appreciate this theory. Price quotes were much lower than the full two-risk factor prOAS model suggests they should have been. According to the APT, such mispricing should allow construction of a flilly hedged risk-free portfolio that earns more than the risk-tree rate. After analyzing a number of trading days, we believe that the TBA-IO dislocation coincides with a sharp surge of rates when the IO market, driven predominandy by the acceleration fear, misses the hedging aspect against slowing down housing turnover. In those market conditions, there is a theoretical opportunity to create a dynamically hedged mortgage portfolio that is prepay-neutral and earns
82 I'uhi'AVMENT R I S K - AND OI'TION-ADJUSTKD V A L U A T I O N of- MBS

an excess return over funding rates. Recognize, however, that the market value of this portfolio would remain exposed to the risk of further TBA-IO dislocation. MODERNIZED GREEKS Valuation adjusted for prepayment risk leads to different rate sensitivity from the traditional approach. Intuitively, premium pass-throughs become less rate-sensitive because their risky spread absorbs interest rate moves following the prepay option moneyness. Indeed, any rate drop elevates the refmancing risk and thereby inflates the traditional OAS; any rate rise reduces the risk and compresses the OAS. Smce the discount MBS react inversely, they are more rate-sensitive under the prOAS method than under the constant OAS risk assessment. A flat OAS profile for the current-coupon to cuspy-premium TBA sector seen in Exhibit 2 suggests that the constant OAS valuation is a valid way to assess their rate sensitivity. All these findings (confirmed in Exhibit 6) can be explained even more easily by the equivalent transition
SUMMEK 20(15

EXHIBIT
Value, % points

5
OAS, bps 2000

Trust IO Valuation as of August 29, 2003

6.0

6.5

7.0
market quotes derived with prOAS; refi risk only 'derived with prOAS; full risk

7.5

S.O

1500

1000

500

-500

zzrzzzzxrzzx

XZZZZZrZZTTZXZXXZZZZZZXZZXZXXXZXZZXZXZ

from the objective to the risk-neutral prepayment model, faster for premiums and slower for discounts. Exhibit 7 compares valuation profiles for an MSR stripped off a 6.5% near current-coupon pool. We see that a constant OAS valuation systematically understates rate sensitivity for all rate levels by as much as one-third. This implies, for example, that MSR managers would underhedge if they used a traditional constant OAS duration. CONCLUDING REMARKS The two-risk tactor prOAS valuation approach that we analyze is a well-defined extension oi the traditional OAS method that draws its roots from arbitrage pricing theory. It successfully explains many phenomena in the MBS market such as OAS variability among MBS coupons and instrument types, the IO-PO pricing paradox, and the divergence of practical durations from the theoretical. At the satne time, the method points to some lacks in the mortgage market that reveal inefficiencies and possible arbitrage.
SUMMER 2O(J5

Two particular anomaliesmissed hedging pow^er of IOs against turnover risk, and exaggerated dynamics of risk priceslet savvy investors construct prepay riskneutral MBS portfohos that earn excess returns and conscientiously speculate on taking a risky position. ENDNOTES The authors thank Jay Delong for help in integrating the model into their valuation system; Dan Szakallas for tuning and optimizing the prepay model to historical prepay data; William Searle for model implementation; and Ilda Pozhegu for publication help. This article has benefited from the comments of Lily Chu, Frank Fabozzi, Yung Lini, Anthony Sanders, and William M. Storms. 'In addition, holding OAS constant for the purpose of computing duration and convexity for au MBS pass-through and its strip derivatives (IO and PO) is inconsistent from a simple mathematical view. When rates change, so do values of the IO and PO, and thus their weights in the pass-through change.

THt JOURNAL <.>F PORTFOLIO MANAGEMENT

83

EXHIBIT 6 From OAS lo prOASTBA Duration Difference A. Market as of May 30,2003 (MTGEFNCL = 4.397%) OAD prOAD Difference FN4.5 4.10 3.42 -0.68 FN5.0 2.90 2.03 -0.83 FN5.5 1.96 1.07 -0.89 FN6.0
1.71

0.77 -0.94

FN6.5 1.67 0.80 -0.87

FN7.0 1.59 0.90 -0.69

FN7.5 1.83 l.ll -0.72

FN8.0 2.12 1.28 -0.84

B. Market as ofAugust FN4.5 OAD 5.78 prOAD 6.11 Difference 0.33 EXHIBIT 7

29, 2003 (MTGEFNCL = 5 .670%) FN5.0 FN5.5 FN6.0 FN6.5 5.18 4.40 3.34 2.48 4.54 5.44 3.32 2.35 0.14 0.26 -0.02 -0.13

FN7.0 2.07 1.98 -0.09

FN'7.5

2.03 1.81 -0.22

FN8.5 1.90 1.68 -0.22

Valuation of 6.5% GWAC MSR as of August 29, 2003

Valuation Multiple
Const OAS Const prOAS

Duration, yrs
10 ^^^^

""*f '
DL ratior
b

/
0 1
k

(f

cu rent :oupc n = 5 67%

- -

Mi Itipie

i-

0.5

1.5

2.5

3.5

4 4.5 5 5.5 6 6.5 Mortgage Market Rate, %

7.5

B.5

9.5

10

-Tbe mixed derivative term appears m the model even if risk factors arc assumed independent. 'Arguably, TBAs should trade even ncber than the agency debt curve because tbcy 1) have superior liquidity, and 2) are collateralized, leading to a perception of higber credit quality. Wbile a typical MBS is backed by properties, tbere is no formal legal mechanism tbat provides any enhanced protection of an agency MBS beyond the corporate guarantee of tbe agency. ""To account for the liquidity difference between IOs and TBAs, we apply 25 basis points prOAS.

REFERENCES

Cheyette, O. "Implied Prepayments." Tlic JournaJ of Portfolio Management, Fall 1996, pp. 107-115. Cohler, G., M. Feldman, and B. Lancaster. "Price of Risk Constant (PORC)." The Journal of Fixed Income, March 1997, pp. 6-15. Davidson, A. "Understanding Prciniuin Mortgage-Backed Securities: Observations and Analysis." In F. Fabozzi, ed., Morrgaj^e-Backed Securities: New Strate_i>ies, Applications and Research.

Chicago: Probus Publishing, 1987. pp. 191-204.

84

PllEHAYMENT RiSK- ftNl) Of I"H)N-Al)JUS"rEI) VALUAI'ION Oh MBS

SUMMER 2005

Davidson, A., M. Herskovitz, and L. Van Drunen. "The Refinancing Threshold Model: An Economic Approach to Valuing MBS." Journal of Real Estate Finance and Econoinics, 1 (June 1988), pp. 117-130. Fabozzi, F., and G. Fong. Advanced Fixed Income Portfolio Mana<iemait. Burr Ridge, IL: Irwin Professional Publishing, 1994. Gahaix, X., A. Krisbnamurthy, and O. Vigneron. "Limit of Arbitrage: Theory and Evidence from the Mortgage-Backed Securities Market." Working paper, 2004. Hayre, L. "Anatomy of Prepayments." Ttie Joumai of Fixed !ncome,]une 2000, pp. 19-49. . "A Simple Statistical Framework for Modehng Bumout and Refinancing Behavior.*' Tliejoutfialof Fixed Income, December 1994, pp. 69-74. Hull, J. Options, Futures, and Other Deriuatii>es, 4th ed. Upper Saddle River, NJ: Prentice-Hall, 2000. Kalotay, A., D. Yang, and F. Fabozzi. "An Option-Theoretic Prepayment Model for Mortgages and Mortgage-Backed Securities." Intemational journal of Theoretical and Applied Finance, Vol. 7, No. 8 (2004), pp. 1-29. Kupiec, P., and A. Kah. "On the Origin and Interpretation of OAS." Tiie journal of Fixed Income, December 1999, pp. 82-92. Levin, A. "Active-Passive Decomposition in Burnout Modeling." Tlxc journal of Fixed Income, March 2001, pp. 27-40. . "Deriving Closed-Form Solutions for Gaussian Pricing Models: A Systematic Time-Domain Approach." International journal of Theoretical and Applied Finance, 1(3) (1998), pp. 348376. . "Divide and Conquer: Exploring New OAS Horizons." Andrew Davidson & Co., 2004.

To order reprints of this article, please contact Ajani Malik at anialik(^iijournab.com or 212-224-3205.

SUMMtK 2(

THEJOUKNAI OF l-*oRTFoir(i MANAGEMENT

85