Looking at the results of the Andrew Davidson & Co. Prepayment Model v4.3.4a for May, we can see that the model was overall a bit slow for FNMA 30s and GNMA 15s and a bit fast for GNMA 30s and FNMA 15s. Across model types and coupons, the model was slightly faster than the actual prepayment speeds for FNMA 15s by a balance weighted average of only 0.3 CPR across coupons. For GNMA 30s the model was faster by a balance weighted average of 1.3 CPR, which was mostly due to 5.5s and 6.0s. The tuning recommendations made a few months ago have greatly helped model fit. For FNMA 30s, the model was slow by a balance-weighted average of -1.4 CPR. For the GNMA 15-year model, the average difference was -1.7 CPR across coupon buckets.
Shifting to the new version 5.1c prepayment model, we can see that the model performed extremely well across all collateral types.
|
FNMA 15 |
FNMA 30 |
GNMA 15 |
GNMA 30 |
Coupon |
Actual |
Model |
Actual |
Model |
Actual |
Model |
Actual |
Model |
4.5 |
10.4 |
10.2 |
8.4 |
8.2 |
13.8 |
11.0 |
14.9 |
12.9 |
5.0 |
13.6 |
12.8 |
12.5 |
11.7 |
17.0 |
12.4 |
17.7 |
15.8 |
5.5 |
17.2 |
17.3 |
17.8 |
15.1 |
20.1 |
17.4 |
23.6 |
21.4 |
6.0 |
21.2 |
20.7 |
25.6 |
25.3 |
21.4 |
20.2 |
28.6 |
30.9 |
6.5 |
23.7 |
21.8 |
31.9 |
29.8 |
22.4 |
21.3 |
34.2 |
36.8 |
7.0 |
23.8 |
21.8 |
34.3 |
32.1 |
21.2 |
21.4 |
34.8 |
37.4 |
7.5 |
25.0 |
22.1 |
34.2 |
34.0 |
25.2 |
21.2 |
33.9 |
34.5 |
Looking at the different pools, the model performed the best for the GNMA 30-year pools, off by a balance-weighted average of just -0.3, and each coupon bucket was no more than 2.6 CPR different between model and actual. The FNMA 30-year model also performed well, as it was off by only -1.6 CPR across coupons. The 15-year models were just as accurate, as the FNMA 15-year model was slightly slow by a balance-weighted average of just -0.5 CPR. The GNMA 15-yr model was slow by a balance-weighted average of -3.1 CPR across coupons, which was due to slightly faster speeds across the coupon stack. It is very clear that the version 5.1c is a very accurate model, and we highly recommend that users talk to their vendor systems about upgrading to this new model.
>>>
