 |
 |
Model Performance
New Tuning Recommendations
by Dan Szakallas
Prepayments showed small increases for low coupons and cuspy premiums in the month of October. The FNMA 15 and 30-years showed small declines in higher premiums, exhibiting the effects of burnout. GNMA's increased almost across the board, as only the GNMA 15-year 7.5's posted a slight decrease of 5.7% from September. New recommended tunings to v433k for the conventional 15 year prepayment models are discussed below. The table shows actual pool CPR speeds of selected coupon buckets, along with corresponding model forecasts.
|
|
FNMA 15
|
FNMA 30
|
GNMA 15
|
GNMA 30
|
||||
|
Coupon
|
Actual
|
Model
|
Actual
|
Model
|
Actual
|
Model
|
Actual
|
Model
|
|
4.5
|
8.6%
|
15.7%
|
7.1%
|
3.0%
|
11.3%
|
3.1%
|
11.2%
|
3.3%
|
|
5.0
|
13.0%
|
26.3%
|
10.7%
|
7.6%
|
13.5%
|
10.1%
|
15.5%
|
7.9%
|
|
5.5
|
20.2%
|
27.4%
|
18.0%
|
20.6%
|
20.1%
|
19.2%
|
21.6%
|
17.1%
|
|
6.0
|
23.3%
|
26.7%
|
28.4%
|
33.5%
|
23.0%
|
21.8%
|
29.0%
|
29.8%
|
|
6.5
|
23.2%
|
27.3%
|
33.6%
|
39.8%
|
24.3%
|
23.4%
|
38.0%
|
36.7%
|
|
7.0
|
24.9%
|
30.1%
|
36.0%
|
43.3%
|
23.5%
|
23.2%
|
41.1%
|
38.9%
|
|
7.5
|
23.5%
|
32.3%
|
37.1%
|
43.6%
|
23.3%
|
23.0%
|
40.7%
|
39.3%
|
The Andrew Davidson & Co., Inc. Prepayment Model v4.3.3k was slightly slower than the actual prepayment speeds for the GNMA pools, and slightly faster than actuals for the FNMA pools. Across collateral types, the model was off by a balance-weighted average of -3.2 CPR for the GNMA 30-year pools, and a balance-weighted average of -3.5 CPR for the GNMA 15's. The FNMA 30-year model was fast by a balance-weighted average of 2.0 CPR across coupons, the top performing model this month. The FNMA 15-year model continues to perform below expectations.
Reviewing model performance over the last 6 months, we have concluded
that specific tuning can definitely improve fit. Without tuning, the
model is fast by an average of 8.8 CPR across the coupons in the table
above. By changing the slide parameter to .91, and the turnover parameter
to 1.4, the overall error actually becomes 0 CPR. You can see the impact
of these changes by going to http://dynamic.ad-co.com
and entering these tuning values on any of the analysis pages. >>>
