Prepayment Modelling

Jason Vinar
U of MN

Spring 2011

Goals

Our goal is to build a prepayment model that we can use to dynamically in scenario, total return, and dynamic analysis of mortgage backed securities We will rst look at some historical data in the context of drivers of prepayment behaviour This will help us dene a model specication that we will use to t model to actual data In our examination of the historical data we will note several problems in using the historical data directly Lastly, we will proceed to t a prepayment model as a case study in class

The Importance of Prepayment Modelling

Prepayments are the primary distinguishing feature of MBS Understanding and forecasting prepayments is critical to the analysis of MBS There is a very important link in the performance of loans and in turn the performance of the pass though or CMO, i.e. individuals make prepayment decisions Changes in prepayments can drastically change the value of the security Knowledge of historical prepayment drivers is key to forecasting future prepayments Prepayment comes from turnover, default, renancing, and curtailments or partial prepayments

Sources of Prepayment

Moving or Turnover
Some FHA/VA mortgages are assumable, which mitigates turnover Factors that drive this are
job changes marriage, divorce, growing families

Default or Involuntary Prepayment

Technically not a prepayment, however, in agency mortgages the full principal is returned to the investor, i.e. the investor is protected against credit risk Since this component has been small historically it is often not modelled directly
Increasingly it is being separated and estimated using loan level data Drivers are unemployment, divorce, and lost equity

Sources of Prepayment, cont.

Renancing
Largest and most variable component of Prepayments Rate/Term Renancing: In low rate environments borrowers opt to renance to lower their payment Cash-out Renancing: Borrowers nance more than they owe on the existing loan and receive the balance in cash at closing There is a cost (both time and money) associated with this which causes each borrower to behave dierently given the same economic condition

Curtailment is a partial prepayment of a loan and is often not modelled

Prepayment Data

Early models were based purely on the borrowers option and overstated prepayments Quickly researchers noted that borrowers acted ineciently and adapted to model this Pool level data is easy to obtain as it is a by-product of trading activity; increasingly loan level data is becoming available Range of coupons is narrow, e.g. when rates are at 5 Age and coupon are the two most important factors
recall the PSA curve the coupon is compared to the current available rate

Historical Data
Fannie Mae, 30 year, xed rate loans originated in 2000. Since some loans we originated in January, some February, etc. it will be dicult to get a good age ramping eect. Prepayments are separated for 5 pass through rates. We can assume the WAC is 50bps higher than the pass through rate. The 10-year US treasury yield for the current month (and lagged by 2 months) is included. All data goes through December 2009. Recall that the nancial crisis started in late 2007.

Analyzing Prepayments
By close examination of the data we can see the behaviour we expected Renance incentives (RI) relative to current mortgage rates Age ramp, prepayment starts low, increase, and then level o Burnout (BO), remaining loans are not as responsive to low rates Lagged response to interest rate lows Seasonality, people are more likely to move in the summer Given the data we will construct a prepayment model using the following functional form as a guide SMM = SMMAGE SEASONALITY + BO SMMRI

Renance Incentive Calculations

The REFINANCE INCENTIVE (RI ) is a way to measure the money-ness of the borrower in terms of their current mortgage rate to the rate available in the market. We are going to use the 10-year semi-annual yield as a proxy for the current MORTGAGE RATE (M). There is no right way to measure the renance incentive. Below is a list of several approaches. Dierence, RI = WAC M Ratio, RI =
WAC M

Payment change, RI = Cold Cnew where Cold is the borrowers current monthly payment and Cnew is the borrowers new payment given the market interest rates and terms. Economic value, estimate of payment savings less the renancing cost over a certain time period. The last two calculations are used most often in loan level models not pool level models.

Renance Incentive or S-Curve

When rates are low relative to the borrowers mortgage rate they renance

The renance incentive here is using the dierence calculation. It also shows 4 time segments of the data to help examine the eect of the nancial crisis. The arctangent function does a good job of tting the s-shape pattern of the renance incentive curve. This curve is commonly referred to as the S-Curve. SMMRI = 1 + 2 atan(3 + 4 RI )

Additional Prepayment Factors

Age (seasoning), due costs associated with home purchase, borrowers who recently purchased a home are less likely to sell the home in the near term. Like the PSA function we will use an age ramp that starts at 0 (corresponding to WALA = 0) and then ramps up to a long term turnover level (SMMLT ) level at WALA = 30. SMMAGE = SMMLT min WALA ,1 30

Burnout is the dampening of the renance incentive and occurs for various reasons. Available mortgage credit could be low. This happened following the nancial crisis. Another possibility is borrowers who missed initial renance opportunities are likely to miss subsequent ones. They are content with the mortgage rate they have and simply will not renance. We are going to focus on the latter and make burnout be a function of the loan age (WALA). BO = f (WALA)

Additional Prepayment Factors, cont.

Time Lag is the eect that prepayments spike about 2 months after the renance incentive is in the money. This occurs because of the time that it takes to apply for and close the new mortgage loan. To model this, it is common to lag the reference mortgage rate by 2 months. Seasonality, housing turnover is more frequent in the summer months primarily corresponding to the school year. Additionally, in colder climates people prefer to house shop and move in the warmer months. The table below shows seasonality factors by month. Jan Feb Mar Apr May Jun 0.66 0.72 1.01 1.07 1.18 1.22 Jul Aug Sep Oct Nov Dec 1.14 1.19 0.99 1.07 0.92 0.83

Prepayment Models
Investing based on prepayment models is a little like driving while looking through the rear view mirror. It may be hard to stay on the road, but it is better than driving with your eyes closed. Davidson Primarily derived from historical experience no matter how simple or complex Usually they are estimated statistically Dicult to predict behaviour of changing products in a changing world Strive to create a robust and parsimonious model
Robust, good in a variety of conditions Parsimonious, capture the biggest drivers using the fewest parameters More complex models tend to t the historical data extremely well, but perform poorly in forward projections

Given both the nancial crisis, making the prepayment data irrelevant going forward, and the desire to better estimate current prices we are going to use forward projections from a market accepted model to build our model. In this approach, we are assuming that the model provider has accurately accounted for historical prepayments levels that are not likely to occur given the current economic outlook.

Review the Prepayment Model

The prepayment model that we will implement has the following functional form SMMt = SMMLT min t , 1 SF (m) + 1 + 2 atan(3 + RI 4 )) 30

where t is the WALA. The parameters dene the shape of the S-curve. The seasonality factor (SF) will be applied for the given month of the forecast. The burnout parameter is omitted as it is both dicult to estimate and implement. For the purposes of this class it is most important to have the correct renance and age relationships. To nd the parameters for the model we will use fmincon in Matlab. This allows us to constrain the parameters so that 0 SMM 1. In our estimation with we only focus on the renance incentive which gives us the ability to add our own view of long term SMM for turnover. The data used in the analysis has been seasonally adjusted as well so we can ignore that term as well. This leaves SMMt = 1 + 2 atan(3 + RI 4 )

Parameter Constraints
The atan function approaches /2 when x and approaches /2 when x . This implies that 1 + 2 2 1 + 2 2 0 1

Using constraints in Matlab we need to dene A and b such that A b. -1 1 0 0 /2 /2 0 0 0 0 0 0 0 0 0 0 0 1 0 0

Initial Guess
It is good practice to analyze the functional forms for an intelligent initial guess. At the very least it will speed up the optimization routine and likely improve your results.

Initial Guess, cont.

Looking at the S-curve, we can make an informed guess for 5 , the long term SMM from turn over. This occurs for aged pass through securities and is clearly seen when there is no incentive to renance and is just min(SMM). From the work for the constraints we already know that 1 2 and = min(SMM) 2

= max(SMM) 2 Solving these two equations for 1 and 2 we get 1 + 2 1 = and 2 = max(SMM) + min(SMM) 2 max(SMM) min(SMM)

In-Class Exercise

Analyze the prepayment function in the context of the S-curve to nd an initial guess for 3 .

Initial Guess, cont.

We can use the y-intercept to help us nd a guess for 3 , i.e. set RI = 0 and use our initial guesses for 1 and 2 . With a little algebra 3 = tan SMMRI =0 1 2

Lastly, 4 determines the slope of the curve or sensitivity of the borrowers for small changes in rates. Higher values for 4 provide a steeper curve and more sensitivity. A reasonable starting point is 4 = 100. An accurate method to nd 4 is to nd the maximum slope of the S-Curve data and analytically do the same using the 1st and 2nd derivative of the prepayment function.

Fitting the Data to the Model

Since we are only estimating the renance incentive parameters we need to rst remove SMM due to turnover. Recall that turnover is a steady state prepayment level that exists with and without the rate renance incentive. Therefore, we can look at the data to determine a reasonable level.

Adding Age and Seasonality in Future Projections

As a matter of application we want to add the seasonality factor into the model. Given the right data we could estimate parameters like the ones shown earlier. However, many prepayment modellers have done this work and have found the parameters to be pretty consistent over time. Therefore, we can use the factors provided earlier. Lastly, we want to add back our age ramp with the steady state turnover. As we use the model we can vary this parameter based on our view of future home sales. Together with the renance incentive parameters we just estimated our nal model is SMMt = SMMLT min t , 1 SF (m) + 1 + 2 atan(3 + RI 4 )) 30

Homework Problem #3
Create a prepayment model that can be used later in our scenario, total return, and dynamic analysis with the following functional form: t , 1 SF (m) + 1 + 2 atan(3 + RI 4 )) SMMt = SMMLT min 30 Observed data to use in the estimation is on the class website. a. Choose a long term turnover level and net it from the prepayment S-curve b. Provide parameter estimates for and compute the sum of squared errors
N

sse =
i=1

(SMMi SMMi )2

where SMMi is the predicted value and SMMi is the observed data net of long term turnover. c. Plot the observed, observed net of turnover, and predicted SMM. d. Create a Matlab function that returns the predicted SMM using the functional form that contains the age ramp, seasonality factor, and renance incentive. The function should have a signature similar to SMM = f (WALA, SMML T , RI , )

References

Davidson, A., Sanders, A., Wol, L., and Ching, A., (2003), Securitization: Structuring and Investment Analysis, Wiley, Hoboken, New Jersey.