You are on page 1of 12

European Journal of Operational Research 166 (2005) 266277 www.elsevier.

com/locate/dsw

O.R. Applications

When to renance a mortgage: A dynamic programming approach


Pei-Ting Lee
a

a,1

, Donald B. Roseneld

b,*

Market Risk, Global Market, Standard Chartered Bank, 6 Battery Road, #03-00, Singapore 659801, Singapore b Sloan School of Management, MIT E40-319, 77 Mass. Ave., Cambridge, MA 02139, USA Received 9 May 2003; accepted 24 February 2004 Available online 14 May 2004

Abstract When should one renance a mortgage loan? It is one of the most common nance questions in todays world. There have been surprisingly few attempts to answer this question in a structured manner, however. Moreover, the existing guidelines for renancing consist of a short list of very simple rules that have a limited application. This article addresses the question through a dynamic programming model coupled with an analysis of historical interest rates. The analysis reveals a more complex set of rules for an optional renance decisionoftentimes conicting with the conventionally accepted idea that rate dierences must be greater than two percent. 2004 Elsevier B.V. All rights reserved.
Keywords: Banking; Finance; Dynamic programming

1. Introduction Many homeowners consider renancing their mortgage at some point in the long life of the loan. Renancingor taking out a new loan to payo an existing loanallows the borrower to take advantage of lower interest rates. The challenge, however, is deciding if and when renancing is advantageous for the borrower.
* Corresponding author. Tel.: +1-617-253-1064; fax: +1-617253-1462. E-mail addresses: lee.pei-ting@sg.standardchartered.com (P.-T. Lee), donrose@mit.edu (D.B. Roseneld). 1 Formerly MIT Operations Research Center, MIT, Cambridge, MA 02139.

While renancing opportunities can appear in many environments, the most signicant area is in mortgage loans for homeowners, due to the long lifetime of the loan and the large amount of money involved. In the United States, this problem is particularly important as nancing policies usually stipulate xed-rate mortgage loans. That is, the rate of interest is xed for the term of the loan, thus implying that a renancing decision can have a large impact. For example, for a $150,000 loan at 7.5 percent, one will pay $377,561 over 30 years $150,000 (the principal amount)plus another $227,561 in interest. A 2 percent drop in the interest rate can translate to roughly $250 less per monthly payment over the 30 years. Thus, there is a large incentive for homeowners to renance if

0377-2217/$ - see front matter 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.ejor.2004.02.017

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

267

current market rates are signicantly less than the mortgage contract rate. Two issues, however, deter homeowners from renancing even when market rates are below the contract rate. The rst issue is cost. When one renances, one generally will repeat many of the same steps, provide the same information, and encounter the same types of costs that were involved in obtaining the original mortgage. (While one can obtain so-called no-points, no-closing-cost loans, rates are higher and one still has to incur time and eort to renance.) The homeowner will therefore have to evaluate whether the potential savings from the lower interest rate counterbalances the costs of renancing the cost. He or she will need to account for the renance charge and number of payment periods left. If the renance charge is too high, then the potential savings of renancing are oset by the renance cost. The second issue is timing. The homeowner has to decide if a better, lower interest rate may be available later. Even if the market rate is below the mortgage contract rate, the homeowner may choose to wait until the market rate is much lower. At the same time, the homeowner must balance the expectation of lower rates with the savings of renancing earlier in the life of the loan. The purpose of this paper is to use analytic approaches to develop guidelines on the renancing question. What are the tradeos in a lower rate as compared with the costs of renancing the loan? We seek to answer theses questions in two ways. First, we develop a modeling framework to address the issue using dynamic programming. Second, we use the model and a computer implementation to develop some simple tradeo curves to give insights on the impact of the key factors in the decision. There are three key factors. These turn out to be the dierence between the rate of interest in the loan and the potential new interest rate, the cost of the renancing as a fraction of the amount of the loan, and the number of periods remaining on the loan. While the two interest rates both have impact, the dierence is what drives decisions. The modeling approach is presented as the appropriate way to address the problem. The guidelines and tradeos both provide insight and also present an alternative to the suggestions and

conventional wisdom that appear in such media as the popular business press in the US. The guidelines and the dynamic programming model represent the two major contributions of this paper. The problem has been treated in a number of sources. (See, for example, Chen and Ling, 1989; Dickinson and Heuson, 1993; Dillon and Stambaugh, 1984; Gyohannes, 1988; Randle and Johnson, 1996; Rupple et al., 1985; Sidgel, 1984; Waller, 1987; Yang and Maris, 1996.) Conventional wisdom, according to Buch and Rhoda (1999), suggests using the 2-2-2 rule as a criterion for renancing: Renancing may make sense if the interest rate potentially available to you is 2 percent less than you are now paying, if you plan to stay in your home for more than two years, and if the renancing charges do not exceed $2,000. This test provides a rough rule of thumb rather than a rigorous mathematical model to help people make a decision. While one might expect alternatives to this conventional wisdom, more recent treatments only give general guidelines (Simon, 2002). This paper takes the opposite route and presents an analytic, dynamic programming model to nd the optimal renancing strategy. Given an amount owed, a contract interest rate, an alternative interest rate, the number of payment periods left (or a distribution on the number of periods left, as many homeowners will sell homes or otherwise payo the mortgage loan before the original term) and the renance cost, the model calculates the total expected costs for (1) servicing the loan under the current contract for one more period and then following an optimal policy, and for (2) renancing at the next stage and then following an optimal policy. Based on this, we can then determine the optimal decision: to renance or not to renance. (One can also look at this as when to renance the rst time. The actual dynamic program structure allows for multiple renancings.) Section 2 presents the model, Section 3 presents the interest model, Section 4 presents the computational implementation, and Section 5 presents results and guidelines for renance. A number of approaches have addressed the pricing of mortgages and mortgage instruments.

268

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

(See, for example, Deng, 1997; Giliberto and Ling, 1992; Kau et al., 1995, Kirby et al., 1990; Rose, 1992; Zipkin, 1993.) There have been some approaches which use option pricing models to model mortgage price (see, for example, Stanton, 1995). Agarwal et al. (2002) then use this approach to address the renancing issue and state that their approach provides the rst analytically tractable model of optimal mortgage renancing. However, these approaches assume that mortgages terminate according to a random hazard rate and only approximate the xed term of a mortgage. With shorter terms the approximation may be an issue. The dynamic programming approach of this article, based on Li (1999), provides an alternative that may be more appropriate with respect to the xed term of the mortgage, particularly with shorter terms.

future expenses such as taxes and insurance), and charges for obtaining title insurance (insurance that certies that the seller has clear title or ownership to the property) and a survey. Closing costs percentage will vary according to the area of the country; lenders or realtors often provide estimates of closing costs to prospective homebuyers. 2.2. The model Let n number of payment periods left in a loan; we will consider both deterministic and stochastic treatment of n contract mortgage interest rate; whenever renancing, the current market interest rate becomes the contract rate market interest rate; this rate is stochastic and varies over time discount factor loan amount owed closing costs to renance the loan ($2,000 is a typical value for M in the US) Boolean decision variable (0 if non-renance, 1 if renance)

b c K M x

2. Dynamic programming model for renance decision 2.1. Assumption Our model has a number of simplifying assumptions. First, renancing does not aect the remaining term of the loan (e.g. if 10 years remain for the original mortgage, then 10 years remain after renancing, although we will consider early payo on a stochastic basis). In our calculations we use a 15-year term of the loan. Second, the market interest rate is assumed to be the average conventional commitment rate. This allows us to use a readily available database to obtain interest rates for our analysis. Third, the discount factor remains the same for each payment period. Finally, we assume that the various costs and fees for renancing are known and can be aggregated into a single value that represents all closing costs. Closing costs are the one-time expenses (over and above the price of the property) incurred by buyers and sellers in transferring ownership of a property. Closing costs typically include what is called an origination fee, an attorneys fee, taxes, an amount placed in escrow (which in the US, for example, allows the lender to insure payment of

a, b, and K are the state variables. (a, b, K) forms the state space. We shall also dene a payment function. This function will enable us to write the recursive cost function in a more compact form later. Let gn i; K periodic payment for servicing a loan of $K in n periods at interest rate i. 2 gn i; K 1 Ki  n :
1 1i

Consider a stream of cash ows C1 , C2 ; . . . ; Cn to be received at date t 1; 2; . . . ; n. Let r be the appropriate discount rate for C. By value additivity the present value of the cash ow stream is C1 C2 Cn : PV 1 r 1 r2 1 rn

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

269

We next dene the following recursive functions. Let fn a; b; K; x total expected cost for servicing the loan given n periods left and in state (a, b, K), if decision x is chosen and an optimal policy is subsequently followed. fn a; b; K optimal total expected cost for servicing the loan given n periods left and in state (a, b, K). The dynamic programming equations for the mortgage renancing problem for xed n are as follows: f0 a; b; K 0; fn a; b; K; 0 gn a; K c Efn1 a; bn1 ; K1 a gn a; K; fn a; b; K; 1
gn b; K c Efn1 b; bn1 ; K1 b gn b; K M;

With the above formulation, we can solve for the optimal cost and decision at any given stage n recursively. In many situations, however, n may not be xed. While any mortgage may have a xed term, a homeowner for a variety of reasons (besides renancing) may pay the mortgage o earlier. The most common reason is a sale of the home, in which case the proceeds from the sale are used to payo the mortgage. In this situation, n is subject to a probability distribution. One way to think about this is that n now represents the nominal remaining length of the mortgage with the condition that there is a probability Pn (depending on n) that the mortgage can be paid o in any given year. The formulation then is changed to the following, with the assumption that the renancing decision in any period is made after any determination is made as to whether the payo of K is to be made: fn a; b; K min fn a; b; K; x1 Pn Pn K
x0 or 1

2 3

where fn is dened by the previous relationships. 3. Stochastic model for interest rates To implement the dynamic programming model, we need to understand the volatility of interest rates (see, for example, Andersen and Lund, 1997) and the distributions of bn1 ; bn2 ; . . . i.e., the interest rates for all future time periods. 3.1. Historical data and random walk model We rst examine the historical data on mortgage rates (Fig. 1, based on www.freddiemac.
10.00 8.00 Interest rate (%) 6.00 4.00 2.00 0.00 time

fn a; b; K min fn a; b; K; x:
x0 or 1

In Eqs. (1) and (2), the expectation E is taken over the distribution of bn1 . We note that the cost fn has two components: a current cost and an expected future cost. Eq. (1) species the non-renancing case as shown by setting the decision variable X to 0. In this case, we will pay gn a; K as the current cost before going to n 1. The new debt becomes K1 a gn a; K, and the contract interest rate remains as a. Therefore, at next stage (stage n 1) we will be in state (a, bn1 , K1 a gn a; K). The expected future cost of this option will thus be Efn1 a; bn1 ; K1 a gn a; K. Note that the recursion allows for multiple renancings. Eq. (2) species the renancing case as shown by setting the decision variable X to 1. In this case, we will pay a current cost of gn b; K M before going to stage n 1. The new debt becomes K1 b gn b; K, and the new contract rate becomes bn . Thus we will be in state (b, bn1 ; K1 b gn b; K) at stage n 1, and the corresponding expected future cost follows.

Fig. 1. Historical weekly data of 15-year xed-rate mortgage (8/30/19912/18/1999).

270

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

com). As can be seen from the graph, there is signicant variation: Over the 7.5-year period, the dierence between the lowest interest rate and the peak interest is nearly 2.5 percent. It is this uctuation that creates situations wherein borrowers want to renance. For example, consider a person who bought a house and took a loan from the bank in 1993 when the mortgage rate was 7.6 percent. After one year, the mortgage rate dropped to 6.74 percent. By renancing, the homeowner could realize a lower interest rate and a potentially better deal. We assume that interest rates follow a zeromean random walk process where the interest rates bt can be written as the interest rate at the previous time period plus a random noise. While this model is fairly simple, it is not inconsistent with historical data. The equation we use to represent this as a discrete time model can be written as: bt bt1 et ; t 1; 2; . . . This can also be written as: rbt bt bt1 et ; t 1; 2; . . . The assumption of a zero-mean random walk implies that the changes in interest rate, et , are independently and identically distributed, with Eet 0, Varet r2 . Therefore, we can specify bt if we know the distribution of et . Given that we assume that the number of years remaining stays the same, the actual interest rate structure could change. However, the eects of this should be relatively low. Appendix A explores the random walk assumption. 3.2. Estimating the increment distribution In this section, we will try to estimate the distribution of the interest rate increments over a period (t 1, t). We rst need to decide on the duration of the time interval (t 1, t). In accordance with the model, (t 1, t) should be a decision interval. At the beginning of each decision interval, the borrower will review his plan and make a decision on whether to renance or to stay on the current plan. Although data are weekly data, in order to solve the dynamic program eciently, we will make the simplied assumption

that the time interval between decisions to renance is a year. In actual practice, renancing decisions are not often made more frequently. To develop the annual distribution, we rst explore the distribution of weekly uctuations using the collected 7.5 years at data. We then can extrapolate this to an annual distribution. For clarity, we use dierent notations for the yearly increments and the weekly increments. We denote the yearly uctuation in interest rate as yt . yt interest rate increment from year t 1 to year t: We denote the weekly increment as wts . wts interest rate increment from week s 1 to week s; in year t: We next represent the yearly distribution as the following equation: yt
52 X s1

wts :

We next must estimate the distribution for the weekly interest rate increments, wts . Table 1 provides the frequency table of the raw data for wts . The sample mean is )0.01, and the variance is 0.0118. The minimum value is )0.35, and the maximum value is 0.45. We next explored various distributions for t. (Although nite sample mean and variance and independence are sucient for the central limit application that follows.) The logistic distribution showed a slightly better t than a normal
Table 1 Frequency table for weekly increments in interest rate Bin <)0.36 )0.36 to )0.27 )0.27 to )0.18 )0.18 to )0.09 )0.09 to 0 0 to 0.09 0.09 to 0.18 0.18 to 0.27 0.26 to 0.36 0.36 to 0.45 0.45 to 54 Frequency 0 2 11 58 157 100 40 17 2 2 1 Cumulative (%) 0.00 0.51 3.33 18.21 58.46 84.10 94.36 98.72 99.23 99.74 100.00

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

271

distribution. Fig. 2 presents the graph. The best logistic distribution t has parameters a 0:0178, b 0:0599. The resulting equation for the density is f x exa=b =b1 exa=b ; where the mean is 0.0178, the variance is 0.0118. Note that the rst parameter is close to zero and the assumption of zero passes the hypothesis test of Appendix A. By the central limit theorem (we assume the wts are i.i.d and the weekly distribution has nite mean and variance), we can approximate yt as a normal distribution, and given the mean and variance of wts , yt $ Normal52lws ; 52r2 : ws With the random walk assumption, we further assume that lws , the mean of wts , is zero. Hence the mean of yt is also zero. r2 is based on the sample ws variance and estimate from the logistic distribution of 0.0118. Thus the distribution for yt and its mean and variance are yt $ Normal0; 0:613496: We thus assume that the interest rate on any year t, bt , can be estimated as the interest rate on the previous year, bt1 , plus a random increment yt which has a normal distribution with parameters (0, 0.613496) and bt bt1 yt ;
4.6
2

4. Implementation 4.1. Automating the dynamic program solution process Before we determine the best time to renance, we automate the solution process for the dynamic programming problem. A formulation with many stages and a normal distribution for the interest rate increments poses computational challenges. We thus need a computer program to solve larger, more realistic problems quickly and accurately. We developed a C++ program with inputs for the program including (1) (2) (3) (4) a: contract interest rate, b: alternative interest rate, n: number of payments left, and the distribution of interest rates increments.

4.2. Discretization of the increments distribution We entered the distribution of the interest rate increments in discretized form. This transformation allows for computational feasibility. The heart of the program is a recursive cost function. It repeats the calculation of expectation for each scenario and nds the optimal value of the cost function. Specically, we discretize the normal distribution in h (the number of realizations of yt ) bins. The bigger h is the better the approximation on the original normal distribution. The diculty lies in balancing precision and eciency. We examined an extreme simplied case; h 2 i.e. yt takes only two values. With such a simple case, we ran the program with a random walk interest rate model. We also explored a more complex case with h 12. The program in principle can be run with any value of h. 4.2.1. Case of h 2 In this case the increments comprise a symmetric Bernoulli distribution, so  d with p 0:5; rb d with p 0:5:

t 1; 2; 3 . . .

Input

2.3
Logistic

0.0 -3.6 -2.7 -1.8 -0.9

0.0

0.9

1.8

2.7

3.6

4.5

Values in 10-1

Fig. 2. Comparison of ()1.78e)2, 5.99e)2).

input distribution

and logistic

272

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

With this input into the program, in each stage we only have to calculate the expectation value of fn1 by
0:5 fn1 a; bn d; Kn1 0:5 fn1 a; bn d; Kn1 :

4.3. Implementation and stage structure Given the structure of the problem, computational eort is an important consideration. The model includes three continuous variables, the interest rate, the market interest rate, and the amount outstanding. This can lead to a computational eort that can grow exponentially, with additional computing power only marginally helpful. The computational task can be addressed in a number of ways. For example, the values of interest rates can be discretized and we can explore a limited set of outcomes. Our approach is somewhat similar in that we use h 2 in the random walk approximation, although we still consider a continuous range of starting interest rates and loan amounts. This approach allows us to run 14 stages with a simple random walk and takes a few seconds, which is sucient for 15-year mortgages. Actual decisions may require more than 15 stages. Our analysis explores 15-year mortgages. 5. Results and analysis This section presents results on renance decision issues for a range of parameters. There are six parameters (a, b, c, K, n, M) that dene the problem, along with some specication on the probabilities of an early payo. In our analysis, we x some of these parameters, and concentrate on investigating the relationships among the others.

The recursivep then much more ecient. function is We set d to 0:613496 0:7833 so that the rst and second moments of the Bernoulli distribution match those of the underlying normal distribution. 4.2.2. Case of h 12 We discretize the range of this normal distribution from 3r to 3r. Here r is the standard deviation of the tted distribution of yearly increments of interest rates, and the probability inside the range is almost 99 percent. For h 12, use bin widths of 6r=12 0:5r. Note that the discretized values are taken as the center of the bins. For the rst and the last bin, the probability is taken as the area under the curve from 2:5r to innity or minus innity to 2:5r, not from 2:5r to 3r (or 3r to 2:5r). This makes the total probability sum to 1. Given these formulae, we can compute the probabilities of the discretized values directly. The following results are obtained using the r as estimated in Section 3, i.e. r 0:7833. The method of h 12 can be applied for each h > 2. For h 2, we t d to the standard deviation of distribution of yearly interest rate increments.

Bin range 3:0 to 2:5r 2:5 to 2:0r 2:0 to 1:5r 1:5 to 1:0r 1:0 to 0:5r 0:5 to 0:0r 0:0 to 0:5r 0:5 to 1:0r 1:0 to 1:5r 1:5 to 2:0r 2:0 to 2:5r 2:5 to 3:0r

Estimated d d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 d11 d12 )2.15 )1.76 )1.37 )0.98 )0.58 )0.20 0.20 0.58 0.98 1.37 1.76 2.15 P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12

Probability 0.00621 0.01654 0.044057 0.091848 0.149882 0.191462 0.191462 0.149882 0.091848 0.044057 0.01654 0.00621

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

273

We believe choosing a and K (as long as M is expressed as a fraction of K) as xed parameters will not result in any loss of generality in our discussion. We use 6.66 percent, which was the 15-year xed mortgage yearly rate in December 1998, as our contract rate a. We also assume the value of the original loan amount K is $300,000. We rst explore n 3 and then n 14. Fig. 3 shows the chart of optimal cost vs. interest rate increment, the latter being the dierence between contract interest rate and market interest rate, (b a). (Even though the random walk assumes only specic values since n 2, one can still compute optimal returns given any differences.) These two plotted curves correspond to a 6:66 percent and a 6:86 percent. We see the value of the optimal costs for the two curves are dierent, but the shapes are similar. The x-values at which the slope changes for the curves are almost the same. As we will see later, the rst point at which the slope changes corresponds to a turning point for decision: Before that point, the optimal decision is to renance; after that point, the optimal decision is to stay on the current contract. The fact that the two curves have similar x-values for the rst turning point means that regardless of the current contract rate, we need only determine the dierence between contract rate and market rate to make our decision of renancing. In the case of 6.66 percent and 6.86 percent, the break-even points are )0.43998 percent and )0.43952 percent respectively. Note that in these examples, the curves show three separate linear regions. The transition from the rst to second corresponds to the transition from nancing to renancing. The second transi-

tion, as noted in the next section, corresponds to suciently high interest rates such that future renancing appears extremely unlikely. In this example, with three stages, there are two such transition points. Additional stages result in additional piece-wise linear regions. 5.1. How does market interest rate aect optimal cost? We next investigate the relationship between the market interest rate and the optimal cost. This informs us as to how a lower market interest rate can be translated into potential savings. Fig. 4 depicts the relationship between the expected cost vs. (b a). Two curves are drawn, one for the decision to renance (which we call the renance curve) and the other for the decision to keep the current contract (which we call the dont renance now curve). The optimal expected cost for any given interest rate dierence will be the lower of the two curves. These curves are plotted using a xed contract rate a 6:66%, loan K $300; 000, stage n 3, c 0:97, closing cost M $200 and using two bins in the discretization of the interest rate increments. From Fig. 4, we can see the renance curve is increasing. In other words, the expected cost for immediate renancing increases as b a increases. This is very intuitive: If we were to renance under a higher market interest rate, the overall cost will certainly increase. On the other hand, it is interesting that the cost for dont renance now also increases with b a, although in a non-linear fashion. This requires additional discussion. If one has chosen not to renance now, why does the cost still increase with the market rate? Why is the market rate relevant at all if one has already decided to stay on his or her existing contract? The explanation goes as follows. Even though we do not renance now, the option to renance later remains open. (Even though this is the case for n 3, the same argument will

342500 342000 341500


cost ($)

a=6.66% a=6.86%

341000 340500 340000 339500 339000 -1 -0.5 0 0.5 1

Interest Rate Increment (%)

Fig. 3. Comparison cost of dierent contract rates n 3.

274

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277
re-finance cost non refinance cost 347000 346000 345000 cost ($) cost ($) 344000 343000 342000 341000 340000 339000 -1 -0.5 0 0.5 1 interest rate increment(%) 341000 340800 340600 340400 340200 340000 339800 339600 339400 -1 -0.5 0 (%) 0.5 1 optimal cost

Fig. 4. Computational result for n 3.

hold for larger n.) If b a is not too large, there is still a good chance that we may choose to renance in the next few periods. As b a increases, the probability for us to renance later decreases. As a result, the optimal cost increases since we become increasingly less likely to enjoy the future benet of a renance option. When b a becomes suciently large, the probability for us to renance at a later stage disappears. The curve therefore plateaus. The plateau shows the expected cost will not increase any more after the market rate exceeds the contract rate by certain value, which we call rp . The expected cost at the plateau corresponds to the decision of dont renance now, or in the future. The market interest b is suciently high that we are condent that it will not fall low enough to warrant a renance in the lifetime of the loan. This phenomenon occurs partly due to our assumption of a nite increment random walk. We assume that the value by which b can decrease in one period is bounded (limited to one standard deviation of the normal in the case of a simple random walk, and about three standard deviations in the case of a general random walk). Thus, if b starts o too high, it will never approach the value of a before the loan contract expires. Let rr be the value of b a where the two curves intersect. (Note that rr is negative.) When the market interest rate is less than the contract rate by more than jrr j, a borrower will get a positive benet from renancing, and the optimal cost will correspond to the optimal decision:

renance now. When the dierence in interest rate b a is more than rr , the renancing cost will be higher than non-renancing. Though the market rate may be below the contract rate over part of this range, the eect of closing cost will oset the prot we get from lower interest rate, so the optimal decision will be dont renance. rr may be called the break-even interest rates difference, since at this point the cost of renancing will oset with the potential benet the borrower gets from renancing. An astute borrower will choose to renance if b a is smaller than rr but will choose to stay on the existing loan contract if b a is bigger than rr , and thus minimize his cost. As a result, the lower of the two curves will give the optimal expected cost. The resultant optimal cost curve consists of three segments: a steep segment for b a < rr , a relatively gradual segment when for rr < b a < rp , and a plateau segment for rp < b a. Figs. 5 and 6 show the charts of expected cost vs. interest rate dierences with the stages n 3 and n 14. From those charts, we nd rp increases as stage n increases. For the case of n 14, the relationship above rr is non-linear, making a more complex relationship than the case of n 3. 5.2. When should we renance? In this section we will investigate the relationships between the break-even interest rates

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277
optimal cost ($)

275

331000 330500 330000 329500 329000 -0.6 -0.4 -0.2 0 0.2 0.4 0.6

interest rate increment(%)

Fig. 5. Optimal cost vs. interest rate increments n 3.

395000

optimal cost ($)

390000 385000 380000 375000 370000 365000 360000 -2 0 2 4 6 8 10

interest rate increment (%)

Fig. 6. Optimal cost vs. interest rate increments n 14.

dierence rr , and two other parameters: the number of stages n, and the closing cost M. We will express the latter as a fraction of the initial loan K. Fig. 7 shows the relation between the number of stages and the break-even interest rates dierence rr , given various values of M=K. This chart provides general guidelines on when to renance. From the chart, some obvious observations came be made. First, for a xed closing cost, jrr j decreases (and b increases) when the number of payment periods left increases. This can be easily understood. When one has less time to clear a loan, one can only balance the closing cost with a lower market interest rate. Conversely, if there are many payments left to enjoy a lower interest rate, a small dierence in interest rate generates a suciently large total savings to cover the closing cost. Second, when the number of payment periods left is xed, the break-even interest rate dierence jrr j will be higher (and b lower) when the closing cost is higher. Thus, when the closing cost is high, one needs a more signicant market rate drop. Finally, when the jrr j is xed, the closing cost will be higher if more payment periods are left.

This can be viewed as a time value for renancing. When the alternative market rate is known, we have toor are more willing topay more to renance if we have more periods left in the life of the loan. How does the possibility of early prepayment and stochastic n aect these results? While a complete analysis would require programming of the equations with the values of Pn , we can use some properties of the problem to estimate the eects. One approach is to assume an equal probability of payo in any period. Note that in this case the problem is very similar to the original problem with discount factor equal to cP , where P is the probability of payo in any period. Alternatively, we can take appropriate averages from the breakeven curves in Fig. 7 (and the cost curves such as Fig. 6). For example, if one knew that one might payo a 15-year mortgage sometime after 10 years, then one could take a mixture of the 10-to-15 year curves. The approximation is not exact, as a 10-year payo of a 15-year mortgage is not exactly the same as a 15-year mortgage, but the recursive equations are otherwise the same. Thus for the case of an early payo possibility, one might use the appropriate mixtures of the curves in Fig. 7. Fig. 7 also suggests the limitations of the traditional 2-2-2 rule (renance if the alternative interest rate is at least 2 percent lower than the contract rate, there are at least two more payments left, and the closing cost is not more than $2000). From our charts, we can nd a more rigorous rule for renancing. For example, assume a M=K from 1/300 to 1/120, which translates to a closing cost of $1000 to $2500 given our assumed initial loan of K $300,000, and number of stages between 3 and 14. The break-even interest rates dierence jrr j does not reach 0.5 percent. This is far less than the 2 percent in 2-2-2 rule. Apparently, the 2-2-2 rule is rather conservative. Given the relative small rr in our nding, there are often good opportunities for renancing that followers of the 2-2-2 rule would pass up, and borrowers should constantly keep a close watch on the market to capitalize on such opportunities. In addition, we also see that, when the number of stages decreases, jrr j increases rapidly in a nonlinear fashion. Thus for a small number of periods,

276

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277
payment periods left
0 0 -0.1 $1000 Critical interest rate increment (%) -0.2 -0.3 -0.4 -0.5 -0.6 $1500 $2000 $2500 K=$300000 c=0.97 5 10 15

Fig. 7. Breakeven interest rate at dierent closing costs.

one will need a larger dierence to justify renancing. With additional periods, a smaller increment will be needed.
week s (%)

0.50 0.40 0.30 0.20 0.10 0.00 -0.10 -0.20 -0.30 -0.40 -0.50 -0.50 -0.30 -0.10 0.10 0.30 0.50

6. Conclusion The renancing problem can be solved by the computer implementation of our formulation. While the problem is simplied and while there are computational limitations, we see that the approach enables a rigorous view of the renance problem. Furthermore, we see that the optimal decision depends on the rate dierences, closing costs and the remaining term. Additionally, much smaller dierences support the decision to renance compared to those suggested by conventional guidelines.

week s-1 (%)

Fig. 8. Scatter plot of weekly change of mortgage rate.

Appendix A. Exploration of random walk model and independence of increments The scatter diagram in Fig. 8 plots the weekly changes of interest rates, i.e. et on successive weeks. For example the point ()0.03, 0.05) refers to two consecutive weeks in which a 0.03 percent decrease was followed by a 0.05 percent increase in mortgage interest rates. If there had been a systematic tendency for decreases to be followed by increases, the northwest quadrant would have many points, and the

southwest quadrant would contain very few. It is obvious from a glance that there is very little consistency in interest rate movements. To explore independence we ran a chi-squared test by dividing up the two-dimensional space into 25 cells obtained by dividing each dimension into ve segments ()0.5 to )0.2, )0.2 to )0.05, )0.05 to 0.05, 0.5 to 0.2, and 0.2 to 0.5). The p value for this test was 17.7%. Collapsing this to a 3 3 or 4 4 set of cells to reduce the number of small-count cells increased the p value further. While this is not conclusive it shows that the data are not inconsistent with independence. We also analyzed the correlation coecient. While zero correlation is necessary, but not sucient, for independence (it is sucient to normal distribution) it also shows consistency. The measured correlation coecient is q 0:000805.

P.-T. Lee, D.B. Roseneld / European Journal of Operational Research 166 (2005) 266277

277

We then tested the hypothesis that q 0 (that the correlation among interest increments is zero). Given a 5%, we nd that F0:05;1;387 3:84 ) F 0:000251. Thus, we do not reject a hypothesis of a zero correlation coecient. We also did a hypothesis test on the mean of the increment being zero, which is necessary for the random walk assumption. Here, n 390, sample mean Ws 5:92 103 , and sample standard ^ deviation s r 0:108824. The Z value for a ttest is Z 5:92 103 p 1:0743: 0:108824= 390

At a 5% with 389 degrees of freedom, Za=2 1:96 > jZ j 1:07 and we thus do not reject the hypothesis of a zero mean process. References
Agarwal, S., Driscoll, J.C., Laibson, D.I., 2002. When Should Borrowers Renance Their Mortgages? National Bureau of Economic Research, Summer Institute. Andersen, T.G., Lund, J., 1997. Estimating continuous-time stochastic volatility models of short-term interest rate. Journal of Econometrics 77 (2), 343377. Buch, J., Rhoda, K., 1999. Marketing mortgage renancing: Suggestions for enhancing borrower choice. Real Estate Finance Journal 14 (3), 3441. Chen, A.H., Ling, D.C., 1989. Optimal mortgage renancing with stochastic interest-rates. AREUEA Journal 17 (3), 278299. Deng, Y.H., 1997. Mortgage termination: An empirical hazard model with a stochastic term structure. Journal of Real Estate Finance 14 (3), 309331. Dickinson, A., Heuson, A.J., 1993. Explaining renancing decisions using microdata. Journal of American Real Estate and Urban Economics Association 21 (3), 293311. Dillon, G.J., Stambaugh, C.T., 1984. Renancing mortgages. Journal of Accountancy 158 (1), 107114. Giliberto, S.M., Ling, D.C., 1992. An empirical-investigation of the contingent-claims approach to pricing residential mort-

gage debt. Journal of American Real Estate and Urban Economics Association 20 (3), 393426. Gyohannes, A, 1988. Mortgage renancing. Journal of Consumer Aairs 22 (1), 8595. Kau, J.B., Keenan, D.C., Muller, W.J., Epperson, J.F., 1995. The valuation at origination of xed-rate mortgages with default and prepayment. Journal of Real Estate Finance and Economics 11 (1), 536. Kirby, R.O., Nash, R.T., Stanford, J., 1990. Mortgage renancing: A high yield investment. The Real Estate Appraiser & Analyst 56 (3), 48. Li, P.-T., 1999. Dynamic Programming Model for Mortgage Renancing Problem with Stochastic Interest Rates. Masters Thesis, MIT. Randle, P.A., Johnson, I.R., 1996. The mortgage renancing decision: A break-even approach. The CPA Journal 66 (2), 6971. Rose, C.C., 1992. Real estate investment renancing. Real Estate Finance 8 (4), 5762. Rupple, G.L., Shilling, J.D., Sirmans, C.F., 1985. Does it pay to renance an FRM with an ARM? Real Estate Review 15 (3), 7982. Sidgel, J.J., 1984. The mortgage renancing decision. Housing Finance Review 3 (1), 9197. Simon, R., 2002. Damage control: What you need to know. Wall Street Journal, June 27. Stanton, R., 1995. Rational prepayment and the valuation of mortgage backed securities. Review of Financial Studies 8, 676708. Waller, N.G., 1987. The optimal timing of residential mortgage renancing. The Real Estate Appraiser & Analyst 53 (2), 2324. Yang, T.T., Maris, B.A., 1996. Mortgage prepayment with an uncertain holding period. Journal of Real Estate Finance 12 (2), 179194. Zipkin, P., 1993. Mortgages and Markov chains: A simplied evaluation model. Management Science 39 (6), 683 691.

Websites www.fammiemae.com www.freddiemac.com www.homepath.com

You might also like