A Variable-Rate Loan-Prepayment Model for Australian Mortgages

by

John Daniel

Abstract: This paper is an investigation of Australian mortgage-loan prepayment from a modelling perspective. A prepayment model for loans of mortgage-backed securities is developed specifically for the Australian mortgage market, and then empirically tested using (Reuters) Australian mortgage-backed security data. The model has origins in the variable-rate loan-prepayment models of the U.S., but is designed and developed to take into account the Australian mortgage-market structure. The model proves very successful when tested empirically, and is able to explain the partialprepayment features of the Australian market as well as full prepayments. Keywords:
PREPAYMENT MODELLING; AUSTRALIAN AND U.S. MORTGAGE MARKETS; VARIABLE-RATE LOANS; FIXED-RATE LOANS; MORTGAGE-BACKED SECURITIES; MORTGAGE RATES; PARTIAL PREPAYMENT; INVESTMENTS. School of Finance and Applied Statistics, Australian National University. Email: John.Daniel@anu.edu.au
The author would like to acknowledge that the mortgage-backed security prepayment data, used for empirical tests of the models, was provided by SIRCASecurities Industry Research Centre of AsiaPacific. Also comments and suggestions by Tom Smith are acknowledged with thanks.
Australian Journal of Management, Vol. 33, No. 2 December 2008, The University of New South Wales

277

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

1.

Introduction

he development and empirical testing of a mortgage-backed security, (MBS), prepayment model appropriate for the Australian mortgage market is the focus of this paper. The main reason for creating prepayment models is to inform MBS investors about what the determinants of prepayments are, and how changes in these determinants will affect prepayments. Prepayment estimation (achieved through prepayment modelling) is a necessary intermediate step in the process of MBS valuation. Prepayment models are also more broadly relevant to all those associated with mortgage finance. Previous Australian prepayment modelling research has been restricted (by researcher choice) to fixed-rate loan prepayment, following the United States (U.S.) fixed-rate prepayment modelling paradigms. This paper expands Australian prepayment modelling research to include variable-rate loans and partial prepayment. 1 Australian mortgage-holder partial prepayment is recognized in such papers as Westpac (2002) and Deutsche Morgan Grenfell (1998); both these papers offer various explanations for partial prepayment. Adding a partial-prepayment component to prepayment modelling is one of the main research contributions of this paper. Most of the research in the field of MBSs (and in particular MBS prepayment) has occurred in the U.S.; hence throughout the paper the U.S. mortgage market is used as a recurring source for comparisons and modelling prototypes. The prepayment model which is investigated here is an adaptation/augmentation of the variable-rate mortgage models of McConnell and Singh (1991), and Sanyal (1994). Although based on the formerly mentioned models, the model developed and empirically tested here, is new, and incorporates unique Australian features. The data used in the empirical testing of the model (which will be referred to as: the Australian variable-rate mortgage model, or just: the variable-rate mortgage model, when appropriate) revealed that Australian MBS pools are characterized by a wide variation in the percentage of variable-rate mortgages and percentage of fixed-rate mortgages, and are invariably not composed of purely variable-rate mortgages, (VRMs), or purely fixed-rate mortgages,(FRMs). The Australian variable-rate mortgage prepayment model is intended as a generalpurpose prepayment model, and hence does also include independent variables which account for FRM prepayment. Given the statement of the last sentence, the prepayment model of this paper might be more accurately referred to as a combined prepayment model (that is, a prepayment model for variable-rate and fixed-rate prepayment); the variable-rate label can be understood more as

1. For the reader unfamiliar with the terms partial prepayment and full prepayment: Whereas renancing prepayment results in the mortgage holder completely paying out the loan, partial prepayments are payments which exceed the mortgage holders required scheduled monthly payments without paying out the loan completely. Prepayment where the loan is paid out completely is referred to as full prepayment. Hence the division of prepayment into these components can be described by the equation: (total)prepayment = full prepayment + partial prepayment

278

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

describing the origins of the model, rather than the purpose or intended application. 2 1.1 The Main Findings of the Research were: (i) For variable-rate mortgage holders: Both partial prepayment and full prepayment are strongly affected by the amount by which the mortgage holders highest interest-rate level attained over the prior course of the loan exceeds the current market mortgage variable-rate. (ii) The ratio of average after-tax share market returns to mortgage rates was introduced as a variable following from the idea of partial prepayment as an alternative investment to shares. This ratio was found to be highly negatively associated with full prepayments. (iii) The prepayment data revealed that in Australia partial prepayment is on average approximately one third of full prepayments for variable-rate loans, and slightly less (closer to a quarter) of full prepayments for fixed-rate loans. (iv) Partial prepayment is not confined to variable-rate mortgage holders; Australian fixed-rate mortgage holders also choose to partially prepay, as is shown by the next finding. (v) For fixed-rate mortgage holders: There is weak evidence to suggest that the amount by which the fixed-rate mortgage holders interest-rate exceeds the current market mortgage fixedrate, determines their propensity to partially prepay. One finding unrelated to interest-rates was: (vi) (As is the case in the U.S.) the age of the pool (or the average number of months since origination of the mortgages in the pool) is a strong influence on full prepayments. An overview of the paper by sections is as follows: There are, as is expected given the similarities of the U.S. and Australian economies, many similarities between the U.S. and Australian mortgage markets. There are also some important differences; these are discussed in section 2. Section 3 is a description of how the two leading (U.S.) articles on variable-rate loan MBS prepayment modelling: McConnell and Singh (1991), and Sanyal (1994), contribute to the creation of an Australian variable-rate prepayment model. Section 4 presents an empirical test of the variable-rate McConnell and Singh/Sanyal model adapted and augmented as necessary for the Australian mortgage market. Section 5 presents concluding remarks.

2. There was no evidence that the US practice of forming MBS mortgage pools containing exclusively variable-rate mortgages, occurs in Australia; however, as can be seen from table 2, two of the twelve pools used for empirical tests contained over 98% variable-rate mortgages, so 100% variable-rate mortgage pools possibly do occur in Australian MBS pools.

279

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

2.

Australian Mortgage MarketHow Different from U.S.

The essential differences between the Australian and U.S. mortgage markets are discussed in this section as preparation for the adaptation/expansion of U.S.-based variable-rate prepayment models. 2.1 Tax To clarify the differences between Australian and U.S. mortgage holder tax regimes the terms owner/occupier mortgage holder and investment mortgage holder are introduced. An investment mortgage holder is defined as a mortgage holder who is purchasing for rental purposes; the term owner/occupier mortgage holder is self-explanatory. Australian owner/occupier mortgage holders cannot claim interest paid on their mortgage loans as a tax deduction, whereas Australian investment mortgage holders can. All U.S. mortgage holders, both owner/occupier and investment mortgage holders, can claim the interest paid on their mortgage loans as a tax deduction. These differences in tax rules have significant implications for the prepayment behaviour of Australian and U.S. mortgage holders, as will be seen in following sections, and particularly in the text under the heading Partial Prepayment and Tax in section 3.1.4: Partial Prepayers. 2.2 Fixed-Rate Mortgage Loan Term Periods While in the U.S. the interest-rate on a FRM is fixed for the entire loan period (usually 2530 years), in Australia the interest-rate is fixed for sub-periods,(1, 2, 3,..., 15 years) of the full loan term period. Loans with a fixed term longer than 5 years are very rare in Australia. 2.3 VRM Rates: Set by Lender (AUS), Indexed (US) The Australian VRM can be described as a loan to a (home) borrower where the interest-rate on the loan varies according to the discretion of the lender. Competition amongst lenders ensures that this rate will be the sum of the shortterm interest rate (referred to as the cash rate) set by the Reserve Bank of Australia (RBA) plus some margin rate which can vary from lender to lender. In the U.S. VRM interest-rates are adjusted according to indices. Further discussion of these differences (in how the respective VRM rates of Australia and the U.S. are reset), and their implications for prepayment, occurs in section 3.1.2: Refinancers. 2.4 Ratio of Fixed-rate to Variable-Rate Mortgages Campbell and Cocco (2003) showed that over the years 1985 to 2001 in the U.S. the proportion of fixed-rate loans varied approximately between 30% and 90%. Several sources (including: NAB 2002, and, Research Note 2003) document that the proportion of loans which are fixed-rate in Australia varies approximately between ten and twenty percent. Thus the ratio of fixed-rate loans to variable-rate loans in the U.S. is higher than the corresponding ratio in Australia, and usually much higher; for example, most of the time over 80% of U.S. mortgage loans are fixed-rate, whereas in Australia at least 80% of mortgage loans are variable-rate
280

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

(all the time). 3 FRMs and VRMs can show significantly different prepayment behaviour patterns. 2.5 Partial Prepayments(Curtailments) In comparison to U.S. mortgage prepayment, Australian prepayments most unique characteristic is the high level of partial prepayment. In the U.S. (where the bulk of prepayment modelling research has been carried out), partial prepayment is so insignificant a part of prepayment that inclusion in prepayment models is regarded as unnecessary; 4 (as stated in the introduction) the prepayment data used in the empirical tests presented later in the paper has revealed that Australian partial prepayment is on average approximately one third of full prepayments, making inclusion of partial prepayment in Australian prepayment models essential. Reasons for the extent of the difference of levels of partial prepayments for Australia and the U.S., are investigated, appropriate variables are designed, and an Australian partial prepayment model is formulated in section 3. 2.6 Recourse/Nonrecourse Mortgage Loans In many states of the U.S., such as California, if the mortgage holder chooses to default on the loan, then the lender has recourse only to the property that the mortgage holder is purchasing with the loan. In Australia, the lender recovering debt in cases of default, has recourse to the mortgage holders other assets such as shares, other properties, and so on. This difference in mortgage contracts probably accounts for why default on mortgages is comparatively much rarer in Australia than the U.S. 2.7 Subprime/Low Doc Loans These are the respective U.S. and Australian categories of mortgages for mortgage holders with substandard credit ratings; (these loans are higher-risk loans, which usually means the lender charges higher rates). There is not complete agreement about the true percentage of Australian mortgages that are low-document loans; estimates vary between less than 0.5 percent to 15 percent. The number of lowdocument loans has been rapidly declining in the wake of the subprime crisis. As one lender states: the proportion of low-document loans-home loans which
3. Recently the percentage of fixed-rate loans has increased to just over 20%. Fixed-rate loans made up 23:7% of new mortgages last month, according to mortgage broker Australian Finance Group. This was up from a low of 16:5% in July. (Uren, The Australian, November 6, 2007). (There is no evidence to suggest that the statement: Loans with a fixed term longer than 5 years are very rare in Australia; is no longer accurate.) 4. The research of Hayre, Chaudhary and Young (2000), reveals that partial prepayments of U.S. mortgage holders do in fact become significant in the later years of long-lasting loansthey state: Information from mortgage services indicates that the prepayment rates from curtailments are typically very low. They average about 0:5% CPR [conditional prepayment rate], early in the life of an ageing pool . . . but can ramp up sharply to as much as 15%-20% CPR at the end of a mortgage term. Because U.S. MBSs typically have lives much less than, say fifteen years, such late partial prepayments of mortgages, (occurring within a typical U.S. MBS pool), would not begin until after the MBS has been called. Hence such partial prepayments in the U.S. do not invalidate the practice of ignoring partial prepayment in U.S. MBS prepayment models.

281

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

require little or no proof of a borrowers ability to make their repayments-written by the broker had plummeted from 15 percent to an insignificant level somewhere below 2 percent. (Klan, The Australian, September 20-21, 2008). This compares to almost 15 percent of U.S. loans being categorised as subprime. Data for the percentage of low-document loans within each Australian MBS pool was not included as part of the Reuters prepayment data set used for the model testing in this paper 2.8 Caps/Floors There are no Caps or Floors for Australian VRMs as there are for U.S. VRMs.

3.

Creating an Australian Variable-Rate Loan Prepayment Model

The two leading U.S. VRM prepayment models considered for adaptation to the Australian variable-rate mortgage market, are McConnell and Singh (1991), and Sanyal (1994). Sanyals prepayment model develops from the modelling theory as established by McConnell and Singh. Rather than a review of McConnell and Singh and Sanyal, the following focuses on explanation of how their models are adapted and expanded for Australian variable-rate mortgage prepayment modelling. 3.1 Adapting the VRM Literature McConnell and Singhs VRM model (and hence also Sanyals) is based on the idea that, for the purposes of estimating a prepayment function, individuals who prepay their variable-rate mortgages can be classified as falling into one of three categories: 1) relocators; 2) refinancers; and, 3) switchers. These three categories of prepayers are described, along with the variables which represent them in the Australian version (of McConnell and Singh/Sanyal), in the following three respective subsubsections. 3.1.1 Relocators (or Home-Sellers) These mortgage holders are prepaying due to job relocation or changes to family circumstances. In prepayment modelling terminology this type of prepayment is referred to as exogeneous prepayment, that is, prepayment unrelated to changes in interest-rates. Relocator variables: Based on the assumption that relocators are more likely to move during the summer months, a seasonal indicator variable is defined as:
1, if the month is a summer month, SEASONAL = 0 for other months.

A positive correlation between SEASONAL and CPR is expected. Based on the assumption that the quadratic equation: CPR(exogeneous)= b0 + b1 2 expresses the relationship between the average age of the pool, ( ), and exogeneous prepayment, the transformed variable, , is defined by:

282

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

WAS * 100% MAX{WAS}

= AGEPOOL =

if WAS MAX{WAS}, if WAS > MAX{WAS},

100%

where WAS (an acronym for weighted average seasoning) is the average age of the mortgages in the pool, weighted by the value of the mortgages, and where MAX{WAS} is the maximum WAS over all pools. The variables: AGEPOOL and SQR(AGEPL), (= 2 ), when simultaneously present in the prepayment model, are expected to have positive, and negative correlations with CPR, respectively. 3.1.2 Refinancers Refinancers, whether FRM or VRM refinancers, prepay to take advantage of lower market interest-rates; however unlike switchers they remain within their loan type, fixed-rate or variable-rate. Before describing refinancers in more detail some further background information is necessary. Unlike Australian VRMs whose rates are reset almost immediately following an RBA induced rate changefor example,the banks usually reset their interest-rates on mortgages within a week or two of an RBA interest-rate changethe U.S. VRMs are indexed to a rate such as a Treasury rate or a LIBOR rate and there is usually a lag of up to a year between index resets and the mortgage rate resets. On first impression of the relationship between RBA and bank mortgage rates, in comparison to the U.S. VRM, the Australian VRM could be expected to behave almost as a pure floating-rate security. If the index rate for the U.S. VRM drops sharply during the interval between interest-rate reset dates, refinancers will choose to prepay their current VRM to take out another variable-rate loan linked to the new lower index rate. So their prepayment behav iour is similar to refinancers of FRMs. These refinancers would not be expected to be present in Australian MBS pools if they consist only of VRMs because, as was stated above, (bank) mortgage rates are being continuously reset to the RBA rate. Why Australian VRMs do in fact exhibit refinancing behavior is revealed through the empirical investigations in section 4 and discussed specifically in subsection 4.5. Also because Australian MBS pools can be expected to contain FRMs as well as VRMs, the fixed-rate refinancing effect will need to be accounted for in the variable-rate model. Refinancing by variable-rate mortgagors may also be affected by a burnout factor much as refinancers are in the case of fixed-rate loans. Richard and Roll (1989) report that prepayments on fixed-rate loans are a function of the history of interest-rates. They argue that fixed-rate mortgagors are differentially sensitive to declines in the rate on fixed-rate mortgages when making refinancing decisions. The first time that the market coupon rate on fixed-rate mortgages falls below the coupon rate of an existing mortgage, for example, the most sensitive mortgagors in a pool will refinance. That is, the most rate-sensitive fixed-rate mortgage refinancers will burnout of the pool. The second time that the pool is subject to a decline in the current market rate to this same level, prepayments will be lower than during the first interest-rate cycle. Only if the current rate on fixed-rate mortgages falls below its previous low will the next level of rate-sensitive fixedrate refinancers be induced to refinance their loans. By extension a similar effect
283

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

may be present with refinancers in VRM-backed securities, the difference, in the U.S., being that the variable-rate mortgagors are responding to index changes rather than actual rates. Refinancing Variables: The differential between pool origination coupon rate and the current market mortgage rate is a commonly used variable to represent the refinancing incentive in fixed-rate mortgage pools in the U.S.. Hence FXDdifl is defined as: FXDdifl = max{3yrFXD coupon rate at origination current market 3yrFXD , 0} where 3 yrFXD is the three-year fixed mortgage rate. A positive correlation between FXDdifl and CPR is expected. Two variables are used to represent burnout: The first burnout variable functions almost identically to rNewMin as in McConnell and Singh and Sanyal and will therefore also be denoted as rNewMin . The only differences between the McConnell and Singh / Sanyal version of rNewMin and rNewMin for the Australian prepayment model are: (i) the short-term rate, r , is proxied by the 3 -year fixed mortgage rate; and, (ii) the condition over the previous 12 months, becomes (in the months) since origination of the pool. A positive correlation between rNewMin and CPR is expected. The second burnout variable (denoted by BURNOUT) is designed to capture the effect of the reduction of the differential rate effect (on CPR), as the number of months, from the last increase in the rate differential, increases (only when the rate differential is positive). Hence
0, if (rNewMin = 1), BURNOUT = incremented by 1 if (FXDdifl > 0), unchanged otherwise.

A negative correlation between BURNOUT and CPR is expected. 3.1.3 Switchers Switchers prepay to change the loan type, (for example,switching from fixed to variable or vice versa). McConnell and Singh suggest that switchers base their prepayment decisions upon their expectations regarding the future. Two determinants of these expectations are: (i) The fixed mortgage rate: McConnell and Singh hypothesize that every time the fixed-rate loan rate falls to a new minimum (in the time since pool origination) some of the switcher individuals will prepay (in order to switch to a fixed-rate loan). McConnell and Singh use the long-term rate as a proxy for the fixed mortgage rate. The variable is defined as:
1, if the long-term interest-rate reaches a new minimum NLM = since origination of the pool, 0 otherwise.

284

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

(NLM is an acronym for New Long Minimum.) A positive correlation between NLM and CPR is expected. (ii) (The change in) the slope of the yield curve: The relative change in spread over each month, between long- and short-term interest-rates, (denoted by SLYC), is used as a proxy for changes in the slope of the yield curve. Sanyal and McConnell and Singh disagree about how SLYC will affect CPR; therefore the expected sign of the coefficient of SLYC is left as indeterminate. 3.1.4 Partial Prepayers To complete the Australian prepayment model, the three categories of prepayers of the McConnell and Singh/Sanyal model are supplemented by a fourth category: partial prepayers. The essential difference between mortgage prepayments in the U.S. and mortgage prepayments in Australia is partial prepayments. Following is an investigation of partial prepayment in Australia including how and why there is this difference between partial prepayment behaviour of the U.S. and Australia. Thus the theoretical foundations for modelling partial prepayment, as is required for the variable-rate model, are established. In order to model partial prepayment, the various reasons for partial prepayment are considered; these include: tax advantages; partial prepayment by default; and, aversion to debt and future interest-rate rises. Partial-Prepayment and Tax The information regarding differences between Australian and U.S. mortgage holder tax regimes, as presented in section 2: Australian Mortgage Market-How Different from U.S., under the heading: Tax, is taken as a basis for the following discussion of partial prepayment and tax. The main underlying driving force which is respon sible for the difference between U.S. and Australian partial prepayment is that, while interest payments on mortgages are tax deductible in the U.S. (regardless of whether the mortgage holder is an owner/occupier or an investment mortgage holder), mortgage interest is not tax deductible in Australia unless the mortgage holder is an investment mortgage holder. This difference creates the incentive for owner/occupier mortgage holders in Australia to partially prepay. Excess funds used to prepay mortgages effectively earn the mortgage rate as an after-tax interest-rate. In comparison the US mortgage holder effectively has an interest free loan and partial prepayment does not save interest payments. 5 Australian owner/occupier mortgage holders will only be tempted to consider alternative investments (such as shares or government bonds) for their excess funds when the after-tax return on the alternative investments exceeds the current mortgage rate. The following (hypothetical) example is

5. The Australian investment mortgage holder, (as distinct from an owner/occupier), is in the same position as all US mortgage holders in being able to claim interest paid on their mortgage loans as a tax deduction, and would therefore be expected not to partially prepay; (that is if they do prepay, they would be expected only to make full prepayments). However there is, as yet, no direct statistical evidence to substantiate such a proposition.

285

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

designed to illustrate how partial prepayments on Australian mortgages effectively earn the mortgage rate as an after-tax interest-rate. Example: Partial-Prepayment And Tax Let a mortgagor have a marginal rate of tax of 50% . To simplify let the mortgage rate equal the risk-free rate of 10% . The mortgagor might consider: Option 1 Market Investment: (a) Invest in Share Market, (RISKY): Make assumptions on risk premium, say 5% . Hence: Expected return on Market: Marginal Tax Rate on return of 50%: After Tax Return: Invest in Money Market(Goverment Loans), (NOT RISKY): Return on Money Market: Marginal Tax Rate on return of 50%: After Tax Return: 15.0% 7.5% 7.5% 10.0% 5.0% 5.0%

(b)

Option 2

Partial Prepayment (NOT RISKY): Return on Mortgage(Opportunity Cost): 10.0% Marginal Tax Rate on Mortgage 0.0%: 0.0% After Tax Return: 10.0%

So the mortgagors best financial option, (given these assumptions), is partial prepayment. 6 From consideration of example 3.1.4.1 the variable ALTINV is defined:
R AOA * (1 Marginal Tax Rate) 1,0, ALTINV= max SVR

(1)

where R AOA is the average continuous compounded return on the All Ordinaries Accumulation Index (over the previous twelve months), and SVR is an acronym for standard variable rate, the standard variable-rate in Australia, calculated, (by the RBA), by averaging the variable-rates of Australian commercial lenders.7 A positive correlation between ALTINV and CPR(Partial) is expected. Default Partial-Prepayment The following quote from Westpac Bank (2002), explains why and how partial prepayment by default occurs:

6. A comparable example for the U.S. would have a return on mortgage, (opportunity cost), of zero percent for partial prepayment, so the U.S. mortgagor is better off investing in shares or government bonds. 7. Technical Note: Since the Australian marginal rate of tax was, over the sample data period, 42 cents for every $1 earned between $58,000 and $70,000, and 47 cents for every $1 earned over $70,000, the marginal rate of tax used in calculations, in the empirical investigations later in this article, was set, (somewhat arbitrarily), at 45%, (unlike example 3.1.4.1 where the marginal rate of tax is taken as 50%).

286

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

...as interest-rates fall, borrowers have the option to reduce their loan repayments or maintain them at previous levels. If repayments are left unchanged as rates fall, the difference between the old and new repayment level becomes partial prepayment instead of interest repayment. Of course not all borrowers will maintain repayment levels when interest-rates fall, however, on the whole we can expect this to have an impact, especially as the majority of mortgages in Australia are variable-rate and attract little, if any penalties for early repayments. To illustrate this point, consider the simple example of a $100,000 loan with an annual interest-rate of 10% . Total interest payments for 1 year would be $10,000 initially but if interest-rates fall to 9% , then this would fall to $9,000 . If the repayment was maintained at $10,000 however, the difference of $1,000 would become [partial] prepayment. Based on this example a 1% decrease in interest-rates could be expected to see up to a 1% increase in (partial) prepayment rates. (Australian) banks have adopted the practice of encouraging borrowers to maintain repayment levels in a falling interest-rate environment in the way described by the preceding quote. The direct debit method of repayment of loans, (where the lender deducts mortgage payments directly from the borrowers account, and which is the repayment method used for most loans), has facilitated this process (of encouraging borrowers to maintain repayment levels in a falling interest-rate environment). From consideration of the preceding quote the variable: partial prepayment by default, (denoted by ParDFLT), is defined as:
ParDFLT = max t {SVR} currentSVR ,

(2)

where max t denotes the maximum mortgage interest-rate over the period from pool origination 8 until the current month t. The max t in (2) is included to account for increases in rates after pool origination; that is, the borrower paying by direct debit is assumed to increase repayments in response to rate increases as is contractually required, and maintain repayments at the new higher level (even if rates subsequently fall). A positive correlation between ParDFLT and CPR(Partial) is expected. Aversion to Future Rate Rises The volatility variable, (VOL y ), expresses the idea that the variable-rate mortgage holders aversion to (mortgage) debt increases as the risk of interest-rate rises increases, that is, as interest-rate volatility increases. Mortgage holders will be expected to increase partial prepayments in response to increasing rate volatility. Hence:
VOLy = volatility of SVR over previous 12months.

(3)

8.

Pool origination here, refers to that time when the average age of the mortgages in the pool is zero;
calculation of this time is further discussed in subsubsection 4.1.1.

287

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

A positive correlation between VOL y and CPR(Partial) is expected. Changes in the mortgage holders partial-prepayment behaviour in response to rate volatility would best be tested over a long period, probably thirty years or more; the effect on mortgage holders of the high-volatility (of rates) periods (such as the early to mid nineteen-nineties) could then be compared to low-volatility periods. Unfortunately a sample data set spanning such a breadth of time was not available. Similar arguments to those used to validate volatility as a partial prepayment regressor variable could be used to validate the level of interest-rates as a partial prepayment regressor variable. There would also be similar data requirements for testing the latter variable. Also considered (but not implemented in empirical work) was a one-sided risk measure, since the mortgage holder is only worried about rate rises, not falls; an upper partial moment might be more appropriate. Summary of partial prepayment in Australia : The preceding proposals for variables can be summarized by the following functional relationship:
CPR(for partial prepaymentin Australia) = f(ParDFLT,ALTINV,VOLy ).

(4)

Model (4) is expected to comprise the essential core of the Australian prepayment model; that is, the variables: ParDFLT, ALTINV, and VOL y should comprise the main explanation of Australian MBS prepayment. 3.1.5 The Formulation of an Australian Variable-Rate Prepayment Model Table 1 summarizes the variable-rate prepayment modelling variables and how they are structured according to (an augmented version of) McConnell and Singhs prepayer categories. From these variables the Australian variable-rate model is formed and can be expressed in functional form as:
CPR = f( AGEPOOL,SQR(AGEPL), ParDFLT, ALTINV, FXDdifl, (5) BURNOUT , VOLy , SLYC, NLM ,SEASONAL, r NewMin )

4.

Empirical Evaluation of the Australian Variable-Rate Model

Section 4 presents an empirical evaluation of the Australian variable-rate prepayment model. Subsection 4.1 describes the data used for the empirical testing. Firstly organization of the data into variable-rate and fixed-rate pool categories is shown in subsubsection 4.1.1. How the prepayment data is divided into components corresponding to total, full and partial prepayment is shown in subsubsection 4.1.2. Descriptive statistics and correlation matrix for all variables including total, full and partial prepayment are given in subsubsections 4.1.3 and 4.1.4, respectively.

288

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

Table 1 Prepayer Categories and their (Proposed) Corresponding Variables

Table 1 summarizes the (VRM) prepayment modelling variables and how they are structured according to (an augmented version of) McConnell and Singhs prepayer categories. From these variables the LINREG model is formed. Relocators AGEPOOL SQR(AGEPL) SEASONAL Refinancers FXDdifl BURNOUT Switchers Partial Prepayers ParDFLT ALTINV VOL
y

SLYC
NLM

rNewMin

Legend of variables: AGEPOOL = the relative weighted average age of mortgages in the pool, (relative to the highest weighted average age of mortgages, (WAS), over all pools in the data set); SQR(AGEPL) = AGEPOOL squared; ParDFLT = partial prepayment by default; ALTINV = alternative investment; FXDdifl = 3 year fixed differential; rNewMin = a new minimum in the 3 year fixed mortgage rate; BURNOUT = burnout; y VOL = rate volatility; SLYC = change in the slope of the yield curve; NLM = new long minimum; and, SEASONAL = seasonality.

Although in section 3 (Creating an Australian Variable-Rate Loan Prepayment Model) the variables of the models of McConnell and Singh, and Sanyal were discussed, the forms of their respective models were not considered. McConnell and Singh uses a proportional hazards model form, whereas Sanyal uses a linear regression model form. Preliminary empirical testing showed that the results obtained, for both the McConnell and Singh and the Sanyal model, were so close that presentation of only one them is necessary; the Sanyal model is presented here. The variable-rate model is tested with total, partial, and full prepayment as the dependent variable, in sections 4.2, 4.3, and 4.4, respectively. One of the most noteworthy results for these tests is that the ParDFLT1 and ALTINV variables are highly significant as predictors for full prepayments, whereas they were both intended as predict ors for partial prepayments only. Subsection 4.5 investigates explanations/hypotheses appropriate for these results. 4.1 Data The interest-rate data (over the period 19962003) were obtained from the RBA website. The data obtained included: Three-year fixed mortgage rates. (The notation 3yrFXD is used in formulae). Standard variable rate. (The notation SVR is used in formulae). 10-year goverment bond, as proxy for the long-term interest rate. 1-month T-bill, as proxy for the short-term interest rate.
289

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

The latter two data items were used to calculate the SLYC variable (as described in subsection: 3.1.3 Switchers). All Ordinaries Accumulation index data was also obtained from the RBA for the calculation of the ALTINV variable; (see equation 1). The MBS prepayment data (which were obtained from Reuters via SIRCA 9 ) consist of twelve pools of prepayment data ranging over the period June 1998 to September 2003 . Prepayment data and test results are presented throughout the Empirical Evaluation of the Australian Variable-Rate model section, in whole-number percentage form (rather than decimal-fraction form). 4.1.1 Fixed and Variable Rate Pools Data Categorization Columns 1 and 2 of table 2 show respectively: (1) the percent age of fixed-rate loans at the beginning of the data period; and (2) the percent age of fixed-rate loans at the end of the data period, for the twelve pools; (the numbers in the brackets of column 1 are the respective beginning dates for the Reuters data of each pool, and the data for all pools end at March 2003.) Column 3 shows the percent age fall in fixed-rate loans, (column 1 minus column 2), over the data period. As shown in column 4, the four pools corresponding to rows 2 5 were selected as the fixed-rate pools; they were selected as fixed-rate pools on the basis of the percentage of fixed-rate loans at the beginning of the data period. Clearly these are the only pools with over 50 % of the loans being fixed at the beginning of the data period. The remaining pools are the variable-rate pools, on which the Australian variable-rate model is tested. 10 4.1.2 (total)CPR, CPR Full , and CPR Partial In practice the monthly conditional prepayment rate of an MBS pool, which is referred to as the single monthly mortality, (SMM), is calculated as:
SMM =

total prepayments outstanding principal at beginning of month

where: total prepayments=opening monthly outstanding principal, less closing monthly outstanding principal, less scheduled monthly repayments. Then the annualized conditional prepayment rate, ((total)CPR), is given by:
( total)CPR = 1 (1 SMM ). 12

(6)

CPR Full and CPR Partial components of the (total)CPR data are then calculated using the respective formulae:
Full prepayments 12 CPR Full = 1 1 ( ) , Open balance

(7)

9. Securities Industry Research Centre of Asia-Pacific 10. Using such a criterion (to categorize the pools into fixed and variable) may seem somewhat arbitrary since other criteria (such as the fall in the percent age of fixed-rate loans, with the presumption being that the fall is due to refinancing) might also have been used as a basis for categorization of the data.

290

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

and:
Total Full 12 CPR Partial = 1 1 ( ) , Open balance

(8)

where open balance is the aggregated outstanding principal over all mortgages in the pool at the beginning of each month.

Table 2 Change in Percentage of Fixed-Rate Loans in each Pool, and Categorization of Pools into Fixed and Variable
. FIXED % at: 1.03 (6/98) 55.41 (7/98) 56.66 (2/99) 70.25 (12/98) 69.15 (6/99) 25.13 (9/99) 29.7 (4/99) 1.73 (10/98) 14.53 (8/99) 33.14 (7/00) 9.75 (3/00) 12.21 (3/00) FIXED % (at Mar/03): 5.86 33.28 47.84 40.94 54.91 8.97 14.36 2.04 10.83 31.47 11.88 8.35 % CHANGE 4.83 22.13 8.82 29.31 14.24 16.16 15.34 0.31 3.7 1.67 2.13 3.86 Pool Category (Fixed or Variable) Variable Fixed Fixed Fixed Fixed Variable Variable Variable Variable Variable Variable Variable

4.1.3 Descriptive Statistics Table 3 provides the descriptive statistics for the variable-rate data for each of the variable-rate prepayment model variables; the statistics for the full and partial prepayments are included. Mean (total)CPR, CPR Partial , and CPR Full , are approximately 27.02, 6.36 and 22.08 percent respectively. Evidently (from the mean of CPR Partial / CPR Full ), partial prepayments are usually approximately 33.91% of full prepayments.

291

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

Table 3 Summary Statistics of the Variable-Rate Model Variables, (Variable-Rate Data)

Variable CPR CPR Partial CPR Full CPR Partial /CPR Full ParDFLT ALTINV FXDdifl NLM No.Obs 336 336 336 336 336 336 336 336 336 336 336 336 336 Mean 27.02426 6.355 22.087 33.911 1.281845 0.045704 0.571875 0.023810 0.107143 0.235119 3.681548 1.0836 1.414494 Std Dev 7.486218 3.886 7.891 26.852 1.260241 0.129503 0.961011 0.152683 0.309756 0.424706 4.383458 0.666865 1.428279 Min 6.810000 3.587 5.269 28.783 0.000 0.000 0.0000 0.000 0.000 0.000 0.000 0.000 1.660000 Median 27.07000 6.016 22.128 30.087 1.250000 0.000 0.100000 0.0000 0.000 0.000 3.000000 1.019358 1.540000 Max 55.93000 24.013 58.383 205.521 4.450000 0.761522 3.650000 1.000 1.000 1.000 25.000 2.167686 3.710000

rNewMin
SEASONAL BURNOUT y VOL SLYC

Legend of variables: AGEPOOL = the relative weighted average age of mortgages in the pool, (relative to the highest weighted average age of mortgages, (WAS), over all pools in the data set); SQR(AGEPL) = AGEPOOL squared; ParDFLT = partial prepayment by default; NLM = new long minimum; BURNOUT = burnout; SEASONAL = seasonality; ALTINV = alternative investment; FXDdifl = 3 year fixed differential; y VOL = rate volatility; SLYC = change in the slope of the yield curve; and, rNewMin = a new minimum in the 3 year fixed mortgage rate.

4.1.4 Correlation Matrix Table 4 provides the pairwise correlation matrix for the variables (total)CPR, CPR Partial , CPR Full , and all the independent variables (except NLM and SEASONAL, which are not included for reasons of space, and since they have the least correlation with (total)CPR). Clearly the correlation between (total)CPR and CPR Full , (0.849946) is very high, while the correlation between (total)CPR and CPR Partial , (0.324592) is low.

292

Vol. 33, No. 2

Table 4 Pairwise Correlation Coefficients of Variable-Rate Model Variables(for Variable-Rate Loan Data)
Variable CPR Partial CPR Full AGEPOOL SQR(AGEPL) ParDFLT ALTINV FXDdifl CPR 0.324592 0.849946 0.104148 0.500139 0.048650 0.460684 0.070681 0.464493 0.110296 0.183919 0.003032 0.290776 0.094995 0.013309 0.046097 0.162340 0.194605 0.141895 0.046084 0.014544 0.030887 0.504714 0.478471 0.408756 0.196845 0.243122 0.018453 0.137593 0.198229 0.128697 0.975954 0.648268 0.321952 0.392838 0.069217 0.408773 0.340172 0.107451 CPR Partial CPR Full AGEPOOL SQR(AGEPL) ParDFLT ALTINV FXDdifl

rNewMin

BURNOUT

VOL

rNewMin
BURNOUT VOL
y

0.630425 0.297609 0.416709 0.066748 0.411378 0.210038 0.129847

0.153668 0.849569 0.126271 0.304806 0.122502 0.250238

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

0.056497 0.005631

0.173523 0.294340 0.028289 0.086699 0.037746 0.136141 0.137355

293

0.105423 0.355014 0.309673 0.038626 0.153481 0.025773

SLYC

Legend of variables: AGEPOOL = the relative weighted average age of mortgages in the pool, (relative to the highest weighted average age of mortgages, (WAS), over all pools in the data set); SQR(AGEPL) = AGEPOOL squared; ParDFLT = partial prepayment by default; ALTINV = alternative investment; FXDdifl = 3year fixed differential; rNewMin = a new minimum in the 3 year fixed mortgage rate; BURNOUT = burnout; y VOL = rate volatility; SLYC = change in the slope of the yield curve. (SEASONAL and NLM were the variables least correlated with CPR, (for the variable-rate data set), and were therefore, for space considerations, not included in table 4).

293

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

There are strong correlations between the time variables (AGEPOOL and SQR(AGEPL)) and the variables, ParDFLT and ALTINV. Another potential source of multi-collinearity is the strong correlation between ParDFLT and FXDdifl. 11 4.2 (total)CPR as Dependent Variable The testing of the Australian Variable-Rate model (as given by equation (5)) begins with (total)CPR as dependent variable. Equation (5) is repeated here for the readers convenience:
CPR = f( AGEPOOL,SQR(AGEPL), ParDFLT, ALTINV, FXDdifl, BURNOUT , VOLy , SLYC, NLM,SEASONAL, r NewMin )

This is the unrestricted form of the model. The tests show how a more parsimonious, but equally effective, form of the model can be found. Because the Australian Variable-Rate model is essentially a linear regression model, univariate tests of the respective variables reveal which variables have potential significance in restricted forms of the model, and which variables can be immediately eliminated from further consideration. Univariate Tests: The following variables were found to have significant effect on CPR at 5% level (or better) in the univariate tests: AGEPOOL, SQR(AGEPL), ParDFLT1, ALTINV, FXDdifl, VOL y 1 , SLYC1, where the 1 suffixes denote the variables are lagged by one month. The variables SEASONAL, rNewMin , BURNOUT and NLM (and variables derived from their lags) were either not significant or had incorrect signs (for the theory); these variables were therefore eliminated from further tests. The results for the univariate tests are shown in table 5. The variables AGEPOOL, SQR(AGEPL), and ParDFLT1 have the highest t-statistics. Revised Unrestricted Model Tests: These seven variables: AGEPOOL, SQR(AGEPL), ParDFLT1, ALTINV, FXDdifl, VOL y 1 , SLYC1, then formed a revised unrestricted model, which was tested on the data. The results for this test are shown in column 1 of table 6. Applying Liklihood-Ratio Tests and the Bayesian Information Criterion to this latter model (as shown in
11. That ParDFLT and FXDdifl are highly correlated is not very surprising; for example, if the SVR and 3yrFXD rates are equal (observation shows they are invariably very close) and rates are monotonically decreasing, then ParDFLT and FXDdifl are identical. However that is hypothetical since over the data period 19982003, the interest rates were certainly not monotonically decreasing.

294

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

columns 2 and 3 of table 6) resulted in the four variables: ALTINV, FXDdifl, VOL y 1 , and SLYC1, being found redundant, leaving a best restricted Australian variable-rate model for (total)CPR with variables: AGEPOOL, SQR(AGEPL), ParDFLT1, and represented in equation form by:
(total )CPR = const + b1 * AGEPOOL+ b2 *SQR( AGEPL) + b3 *ParDFLT 1

(9)

Attempts at further interpretation of these results are postponed until the results of the corresponding tests with CPR Partial and CPR Full as dependent variable have been completed. 4.3 CPR Partial as Dependent Variable Partial prepayment theory has only been developed for the variables ParDFLT, ALTINV, and VOL y (in subsubsection 3.1.4); hence the expected signs, of the coefficients of all other independent variables for CPR Partial , are indeterminate. Univariate Tests: The univariate tests with CPR Partial as dependent variable (table not shown here) find that only the variable ParDFLT1 is significant at 5% level and to have correct sign. Though having no theory for partial prepayment, the FXDdifl variable is significant at 10% level. That FXDdifl variable is significant at 10% level might be taken as a weak sign that partial prepayment is not exclusive to variable-rate mortgage holders. A positive relationship between partial prepayment and FXDdifl is, on first impression, difficult to comprehend. If market mortgage rates are lower than fixedrate borrowers rates (when FXDdifl is larger), the anxieties of these borrowers about having to pay higher rates in the near future (when their fixed-term expires and their rates are reset to the current market rate) should be less. However possibly any rate changes are, to fixed-rate borrowers, a reminder of the volatility of interest rates, and hence a warning that in a few years time they may be paying higher rates; hence their reaction of partial prepayment. 12 Revised Unrestricted Model Tests: Column 1 with heading (Revised) unrestricted model in table 7 shows the results for a regression with the two variables ParDFLT1 and FXDdifl as independent variables. The sign of the coefficient of FXDdifl has changed to negative, presumably due to multi-collinearity with the ParDFLT1 variable; (correlation(ParDFLT1, FXDdifl) = 0.849569 from table 4). Column 2 (Best (restricted) model) of table 7, shows the results for a Liklihood-Ratio Test of the hypothesis that the variable FXDdifl is redundant to the model of column 1 . Clearly the Liklihood-Ratio Test shows that the hypothesis that FXDdifl is redundant cannot be rejected.
12. In Daniel (2006), in the tests on the combined data, two other fixed-rate variables, namely the two burnout variables: rNewMin and BURNOUT, were also found significant for partial prepayment, reinforcing the evidence for the idea that partial prepayment is not exclusive to variable-rate mortgage holders.

295

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

Table 5 Univariate Variable-Rate Model Variable Coefficients, t-Statistics, pValues and Adjusted R 2 s, (with (total)CPR as Dependent Variable)
Independent Variables: AGEPOOL SQR(AGEPL) ParDFLT1 ALTINV FXDdifl1 BURNOUT1 VOL 1
y

Expected Sign of Coefficient +

Coefficient and t-statistic 0.2031 (10.5553) 0.1908 (9.4859) 2.8136 (9.8286) 1.4761 (3.7497) 2.2425 (5.4935) 0.2920 (3.1710) 2.0622 (3.4153) 0.8820 (3.1196) 2.5200 (2.7832) 0.3257 (0.2432) 5.4694 (2.0515)

p-value 0.0000 0.0000 0.0000 0.0002 0.0000 0.0017 0.0007 0.0020 0.0056 0.8080 0.0410

Adjusted 0.2479 0.2099 0.2236 0.0375 0.0824 0.0263 0.0309 0.0254 0.0173 0.0028 0.0095

R2

+ -+ -+

SLYC1
SEASONAL 1

+ + +

rNewMin
NLM1

Note:

The coefficient values, t-statistics and p-values are for individual variables only, while the adjusted 2 R s are for each of the complete regression models (which include a constant). Because SQR(AGEPL) is not designed to occur as an independent variable without AGEPOOL, the expected sign of the coefficient of SQR(AGEPL) is indeterminate.

Legend of variables: AGEPOOL = the relative weighted average age of mortgages in the pool (relative to the highest weighted average age of mortgages, (WAS) over all pools in the data set); SQR(AGEPL) = AGEPOOL squared; ParDFLT = partial prepayment by default; ALTINV = alternative investment; FXDdifl = 3-year fixed differential; NLM = new long minimum; rNewMin = a new minimum in the 3-year fixed mortgage rate; SEASONAL = seasonality; BURNOUT = burnout; y VOL = rate volatility; and, SLYC = change in the slope of the yield curve.

296

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

Table 6 Variable-Rate Model Coefficients and t-Statistics, (with (total)CPR as Dependent Variable)
(Revised)1 Unrestricted-Model Column 1 (total)CPR No. of Obs Independent Variables: Constant AGEPOOL SQR(AGEPL) ParDFLT1 ALTINV FXDdifl1 VOL 1
y

Restricted-Model Column 2 Dependent Variable: (total)CPR 336 12.4622 (5.1328) 0.4698 (4.1741) 0.3306 (3.0792) 1.4337 (3.9014) 0.0423 (0.0432)

Best-(Restricted) Model Column 3 (total)CPR 336 13.82566 (7.017232) 7.017232 (4.238887) 0.242452 (2.789269) 1.540579 (4.317186)

336 12.08488 (4.290840) 0.426116 (3.200806) 0.282079 (2.307838) 2.143346 (2.429146) 0.348102 (0.739369) 1.158727 (1.203356) 0.874345 (1.267886) 0.249342 (0.798438) 1088.236

0.99624 (1.4236)

SLYC1
Log-Likelihood Likelihood Ratio Statistic p-val BIC2 adjusted R Note:
1 2

6.616095 0.3039

1090.492 4.5126 0.104734 6.5949 0.2988

1091.529 6.586681 0.159410 6.566447 0.2987

Revised means after variables rejected by univariate regressions have been eliminated. BIC denotes the Bayesian Information Criterion.

Legend of variables: AGEPOOL = the relative weighted average age of mortgages in the pool, (relative to the highest weighted average age of mortgages, (WAS), over all pools in the data set); SQR(AGEPL) = AGEPOOL squared; ParDFLT = partial prepayment by default; ALTINV = alternative investment; FXDdifl = 3 -year fixed differential; NLM = new long minimum, SEASONAL is seasonality; rNewMin = a new minimum in the 3 -year fixed mortgage rate; BURNOUT = burnout; y VOL = rate volatility; and, SLYC = change in the slope of the yield curve.

297

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

In this best (restricted) model ParDFLT1 is significant at 5% level and of correct sign. Although the significance level of the best model overall is low (adjusted R 2 = 0.011501 ), both this value and the BIC, ( 5.569520 ), confirm that the best model is an improvement on the unrestricted model with adjusted R 2 = 0.008644 and BIC = 5.586721 . The best restricted Australian variable-rate model for CPR Partial , in equation form, is given as:
CPRPartial = const + b1 *ParDFLT 1

(10)

Neither of our other two partial prepayment variables (the ratio of average after-tax share market returns to mortgage rates, and the volatility of mortgage rates), as proposed in subsubsection 3.1.4, showed a significant impact on variable-rate mortgage holders partial prepayment.

Table 7 Variable-Rate Model Coefficients and t-Statistics, (with CPR Partial as Dependent Variable), (Variable-Rate Data).
(Revised)1 Un-restricted Model Column 1 CPR Partial No. of Obs Independent Variables: Constant ParDFLT1 FXDdifl Log-Likelihood Liklihood Ratio Statistic p-value BIC2 2 adjusted R Note:
1 2

Best (restricted) Model Column 2 CPR Partial 336 5.879847 (19.54441) 0.370654 (2.213088)

Dependent Variable:

336 5.860990 (18.50913) 0.420482 (1.368093) 0.077862 (0.193468) 929.8435

5.586721 0.008644

929.8623 0.037765 0.845916 5.569520 0.011501

Revised means after variables rejected by univariate regressions have been eliminated. BIC denotes the Bayesian Information Criterion.

Legend of variables: ParDFLT = partial prepayment by default; and, FXDdifl = 3 -year fixed differential.

298

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

4.4 CPR Full as Dependent Variable Univariate Tests: Because the partial prepayment variables have no theory for full prepayment, the expected sign of their coefficients is prima facie indeterminate when CPR Full is the dependent variable. For this reason ParDFLT1, ALTINV and VOL y variables have an indeterminate expected sign in the univariate regression tests (table not shown here). Taking into account the preceding note, the univariate tests, with CPR Full as dependent variable, select the same seven, significant at 5% level and correct sign variables, as the univariate tests with (total)CPR as dependent variable; namely : AGEPOOL, SQR(AGEPL), ParDFLT1, ALTINV, FXDdifl, VOL y 1 , SLYC1. That all three proposed partial prepayment variables, ParDFLT1, ALTINV and VOL y 1 , are significant for full prepayments is noteworthy. The significance of SLYC1 presumably shows evidence of switching from variable-rate to fixed-rate. The positive sign for the coefficient of SLYC1 agrees with the empirical findings of McConnell and Singh rather than Sanyal. Revised Unrestricted Model Tests: The seven variables: AGEPOOL, SQR(AGEPL), ParDFLT1, ALTINV, FXDdifl, VOL y 1 , SLYC1, as selected by the univariate tests, become the variables for the revised unrestricted model, as shown in column 1 of table 8. Applying Liklihood-Ratio Tests (and the Bayesian Information Criterion), again leads to the best restricted model (as shown in column 4 of table 8), with variables AGEPOOL and ParDFLT1. On the basis of the adjusted R 2 s, there is very little difference between the explanatory power of the respective models in table 8. However the Bayesian Information Criterion does show a reduction in value (that is, an improvement in explanatory power) as the models progress from unrestricted to best restricted model. The best restricted Australian variable-rate model for CPR Full , in equation form, is given as:
CPRFull = const + b1 * AGEPOOL+ b2 *ParDFLT1

(11)

4.5 ParDFLT1 and ALTINV Variables Reviewed as Full Prepayment Variables The univariate tests (with CPR Full as dependent variable, in subsection 4.4) showed that ALTINV variable is highly significant, (p-value = 0.0003 ), and negatively associated with full prepayment. 13 Both the univariate and the multivariate tests of subsection 4.4 showed that ParDFLT1 variable is also highly significant as a
13. In Daniel (2006) the empirical testing data, as well as being organised into variable-rate data, was also organised into fixed-rate and combined data sets. The highly significant and negative association between full prepayment and ALTINV was evident in these data sets as well.

299

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

Table 8 Variable-Rate Model Coefficients and t-Statistics, (with CPR Full as Dependent Variable), (Variable-Rate Data)
(Revised)1 Unrestricted Model Column Restricted Model Column Restricted Model Column 3 CPR Full 336 9.925397 (4.670235) 0.300318 3.215028 0.131680 (1.404415) 0.945721 (2.456919) Best (Restricted) Model Column

Dependent Variable: CPR Full No. of Obs Independent Variables: Constant AGEPOOL SQR(AGEPL) ParDFLT1 VOL 1 ALTINV FXDdifl
y

CPR Full 336 9.860966 (4.067029) 0.335559 (2.759425) 0.167117 (1.454507) 0.922059 (2.337623) 0.485149 (0.637013) 2.948770 (0.956808)

CPR Full 336 12.49372 (11.52214) 0.174448 (6.616995)

336 9.635591 (3.512665) 0.320087 (2.413582) 0.159164 (1.301417) 1.113415 (1.259093) 0.336575 (0.440806) 3.646505 (1.166233) 0.412077 (0.433504) 0.439982 (1.325612) 1114.473

0.940813 2.440707

SLYC1
Log-Likelihood L.R.S.2 p-value BIC3 2 adjusted R Note:
1 2

6.772268 0.267612

1116.391 3.837924 0.146759 6.749064 0.263688

1116.974 5.003108 0.286979 6.717907 0.265581

1117.969 6.993346 0.221136 6.706517 0.263437

Revised means after variables rejected by univariate regressions have been eliminated. L.R.S. denotes Liklihood Ratio Statistic. 3 BIC denotes the Bayesian Information Criterion.

Legend of variables: AGEPOOL = the relative weighted average age of mortgages in the pool, (relative to the highest weighted average age of mortgages, (WAS), over all pools in the data set); SQR(AGEPL) = AGEPOOL squared; ParDFLT = partial prepayment by default; ALTINV = alternative investment; y FXDdifl = 3-year fixed differential, VOL is rate volatility; NLM = new long minimum; rNewMin = a new minimum in the 3-year fixed mortgage rate; SEASONAL = seasonality, BURNOUT is burnout; and, SLYC = change in the slope of the yield curve.

300

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

predictor for full prepayments, Both these variables were intended as predictors for partial prepayments only. Subsection 4.5 investigates explanations/hypotheses appropriate for these results. 4.5.1 ParDFLT1 as a Full Prepayment Variable The result ParDFLT1 having strong significance in relation to explaining the variation in CPR Full is inconsistent with our (default partial prepayment) theory (as presented in section 3.1.4: Partial Prepayers, under the heading: Default Partial Prepayment), which proposed ParDFLT as a partial prepayment variable only. As is evident from equation (2) (which defines ParDFLT), the ParDFLT variable is calculated from the variablerate, not the fixed-rate. Full prepayments by variable-rate borrowers (which the empirical test results imply are occuring to a significant extent) can be explained as evidence of: (i) variable-rate refinancing; or, (ii ) switching from variable-rate to fixed-rate loans. Each of these possibilities is now considered.
(i ) Variable-Rate Refinancing: One of the assumptions implicit in our theory prior to empirical testing has been that there is not enough difference between variable-rate lender rates to make a variable-rate borrower fully prepay their existing loan in order to take out another lower-rate variable-rate loan. That is, we have made the assumption that there is no refinancing from variable-rate loan to variable-rate loan. There was some evidence in the literature of variable-rate refinancing which we chose to regard as unimportant. For example, Westpac bank (2002), states: A bigger impact of falling interest-rates [than partial prepayments], is the associated potential increase in full prepayments via refinancing. Changing interest-rate environments tend to create differences between competing lenders rates and so tend to act as a trigger for borrowers to reassess their debt position. As loan products have become readily available with lower entry and exit costs, borrowers have become increasingly sophisticated and refinancing has become more and more commonplace. Borrowers quickly take the opportunity to swap to improved products or cheaper rates when differentials arise leading to an increase in full prepayments.

ParDFLT is a measure of the differential rate between rates of a (variable-rate) borrower repaying by direct debit and the current market variable-rate. In review the possibility that some variable-rate borrowers find refinancing to another more competitive variable-rate loan clearly cannot be excluded.
(ii ) Switching from Variable-Rate to Fixed-Rate Loans: As already mentioned in subsection 4.4, the significance of SLYC as a predictor for full prepayments implies that switching from variable-rate loans to fixed-rate loans is probably occurring. In conclusion (of this review of ParDFLT), we have to remain open about whether these full prepayments of variable-rate loans are just refinancings to other variable-rate loans, or just switches to fixed-rate loans, or some combination of these two modes of full prepayment.

301

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

Whereas previously all variables have represented only one type of prepayer exclusively, now we have a variable which could possibly be representing three types of prepayers. ParDFLT possibly simultaneously represents: (i) partial prepayers; (ii) refinancers (variable-rate to variable-rate); and, (iii) switchers (variable-rate to fixed-rate). Our aim, of creating a model which is as parsimonious as possible, is well served by the multi-faceted nature of the variable ParDFLT. 14 4.5.2 ALTINV as a Full Prepayment Variable We have noted previously (at the beginning of this section) that our proposed partial prepayment variable ALTINV, as a predictor for full prepayments, is univariately significant and has negative sign (and that this result also occurred in the univariate tests for both fixed-rate and combined data, as shown in Daniel 2006). In preparation for the following discussion in relation to the ALTINV variable we note: (i) that (as previously mentioned in section: 3.1.4 Partial prepayers) the Australian investor mortgage holder is in the same position as the US mortgage holder in being able to claim interest paid on their mortgage loans as a tax deduction, and would therefore be expected not to partially prepay; (that is, if they do prepay, they would be expected only to make full prepayments). (ii) that the variation in ALTINV is mainly determined by the variation in the (All Ordinaries) share market index. The basis of the definition of the ALTINV variable (see equation (1)) is the ratio of average after-tax share market return over the previous year to the current mortgage variable rate. Because the mortgage rate is relatively static and the average share-market return more volatile, the variation in ALTINV is mainly determined by the variation in the share-market index. Possibly the ALTINV variable is showing a negative association between the average share-market returns and the willingness of professional investor mortgage holders to sell their properties; alternatively stated, the professional investor mortgage holder recognizes that a time when the share market is booming is not a good time to sell because then shares are a competing investment for potential buyers. If the preceding explanation is correct, the professional investor mortgage holder will also be watching property returns. The best time to sell an investment property would be when property returns are high relative to share returns; the worst time to sell would be when share returns are high relative to property returns. A robustness check on the ALTINV variable was made to check our former surmise(s).

14. The preceding discussion might be regarded as an argument which implies that the name ParDFLT is not entirely appropriate for this variable, which is now recognized as not just a partial prepayment (by default) variable .

302

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

Using the listed property index GICS22, a variable: R AOA R GICS22 (the difference between average return on shares and average return on property) was formed and tests, with CPR Full as dependent variable, carried out. 15 Although univariately very significant and correct sign, in multiple tests the variable became in significant and wrong sign; correlation testing showed these problems were due to multi-collinearity with the time variables, AGEPOOL and SQR(AGEPL). While the latter results do not corroborate our former surmised hypothesis/explanation for ALTINV in relation to full prepayments, we need to take into account the fact that ALTINV is a complex variable. ALTINV does not just reflect share-market index variation; there are several factors in the definition/ calculation of ALTINV which might account for the lack of corroboration by the robustness tests. 4.6 Summary of Variable-Rate Model Results The most noteworthy result of the empirical tests is how well the restricted prepayment model:
CPR = f(AGEPOOL,SQR(AGEPL),ParDFLT),

(12)

performs relative to the unrestricted model (5). The empirical tests also found that the best variable-rate models for partial and full prepayments of Australian MBSs are, respectively:
CPRPartial = const + b1 *ParDFLT 1,

and
CPRFull = const + b1 * AGEPOOL+ b2 *ParDFLT 1

The prepayment data revealed that in Australia partial prepayment is on average approximately one third of full prepayments for variable-rate loans. The ParDFLT variable (representing the amount by which the mortgage holders highest interest-rate level attained over the prior course of the loan, exceeds the current market mortgage variable rate) was found to have a highly significant effect on both partial prepayments and full prepayments for the variable-rate mortgage holder. The ALTINV variable (representing the ratio of average after-tax share market returns to mortgage rates) was found to have a high negative association with variable-rate mortgage holder full prepayments in the univariate tests. (However ALTINVs significance was not great enough to register as a best model variable (in the multivariate tests)). Other univariately significant and correct sign interest-rate variables, for full prepayments of variable-rate mortgage holders, were FXDdifl, VOL y 1 , and SLYC1.
15. The notation here for R AOA and R GICS22 is consistent with the notation for equation (1), (which defines ALTINV).

303

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

That FXDdifl is (weakly) significant for CPR Partial tends to imply that partial prepayment is not confined to variable-rate borrowers, and that partial prepayment varies proportionately with the differential between the fixed-rate mortgage holders interest rate and the current market mortgage fixed rate. The empirical tests of our model(s) have revealed that our independent variables explain full prepayment much more effectively than partial prepayment. This conclusion is reached through comparing the relatively low adjusted R 2 s, 0.008644 and 0.011501 for the unrestricted and restricted models respectively, with CPR Partial as dependent variable (as shown in table 7) with the corresponding adjusted R 2 s for the unrestricted and restricted models with CPR Full as dependent variable, 0.267612 and 0.263437 respectively (as shown in table 8). In Daniel (2006), using the combined data set, more independent variables are found significant as predictors for partial prepayment (namely VOL y , rNewMin and BURNOUT); however the overall significance of the prepayment model(s) for partial prepayment ( 0.057220 and 0.046583 for the unrestricted and restricted models respectively) are only a small improvement on the variable-rate data tests results as shown here.

5.

Conclusion

The objective of this paper has been to investigate Australian mortgage prepayment by developing and testing prepayment models for loans of Australian MBSs. While the foundations for the new Australian prepayment model developed in this paper are in U.S. MBS variable-rate loan prepayment models (notably McConnell and Singh 1991; and Sanyal 1994), the model is structured to account for the differences in the Australian mortgage market (for example, the ramifications of the different tax regimes for the mortgage holder as an investor). The new Australian prepayment model proved successful when tested on (Reuters) Australian MBS data; parsimonious forms of the model were able to successfully explain both total prepayment and the components of total prepayment: full and partial prepayment. (Date of receipt of final transcript: February 26, 2008. Accepted by David Gallagher & Garry Twite, Area Editors.)

References
Campbell, J. & Cocco, J. 2003, Household risk mnagement and optimal mortgage choice, The Quarterly Journal of Economics, vol. 118, pp. 144994. Daniel, J. 2006 Prepayment modelling of (Australian) mortgage loans, Ph.D.thesis, Australian National University. Deutsche Morgan Grenfell, 1998, Mortgage-Backed and Asset-Backed Security Research,, July15, 1998. Klan, A. 2008, Lowdoc loans dead: Symond, The Australian, September 201, 2008, p. 6. Hayre, L., Chaudhary, S. & Young. R. 2000, Anatomy of prepayments, The Journal of Fixed Income June, vol. 10, no. 1, pp. 1949.

304

Vol. 33, No. 2

Daniel: VARIABLE-RATE LOAN-PREPAYMENT MODEL

McConnell, J. & Singh, M. 1991, Prepayments and the valuation of adjustable-rate mortgagebacked securities, The Journal of Fixed Income, vol. 44 , June, pp. 2135. National Australia Bank, 2002, A Guide to Australian MBS, Mortgage and Asset Backed Research, NAB website: www.nabmarkets.com March. Research Note, Department of the Parliamentary Library, 2003, Tightening the Mortgage Belt, no. 21, 24 November, pp. 20034. Richard, S. & Roll, R. 1989, Prepayments on fixed-rate mortgage-backed securities, Journal of Portfolio Management, vol. 15, Spring, pp. 7381. Sanyal, A. 1994, Ammunition for ARMs: A panel data approach to prepayment modelling, Journal of Fixed Income, vol. 4, no. 3, December, pp. 96103. Uren, D. 2007, Final straw in case for an increase, The Australian, November 6, 2007, Business, p 23. Waters, E. 2002, MBS prepayment review, in The Debt Monitor, Westpac Bank. Westpac Bank, 2002, The Debt Monitor.

305

AUSTRALIAN JOURNAL OF MANAGEMENT

December 2008

306