27 March 2015 US Fixed Income Weekly Figure 8: High-LTV pool pay-ups show curved relationship to refi incentive 6.0 5.0 4.0 3.0 00 0.5 Reft Incentive (%) 1.0 1.5 2.0 —Smooth 0 Payup vs Reel Nola ReII menhirs is measured es the wow WAC on the TBA %Went% Ice a given period less the saved% 30-veer primary nwngege rate Sant Deutscht area (Figure 9: Model fit for CO 3.5%s in and out of sample [Figure 10: Model fit for CO 4.0%s In and out of sample 100 250 200 150 100 030 woo 030 401 1.50 .103 430 ni a SLO 4W 3W 2.00 140 0.03 .103 -2-03 1•1 MI ell fit .% ea 2 2 RR R RR .R RRR R R BR -4 R R R RR aaaaaa 3 3 aaaaaa aassts5ssstsEsas ccs s —ku jer y,sampre —howl lei Wick Oat Steeple --- Fitted Out San* San Deutsche bee son Oftesthe Sint —Actual in5ample —Wed In Swale ActuelOut Semple --- Rao Cki Sample Given this relationship we chose to use a non-parametric general additive model (GAM) to describe that function. Reassuringly, the core function relating ref i incentive to pay-up continues to hold well in out-of-the-money high-LTV pools, which are non-deliverable and so can trade back of TBA. That model produced a fit with nearly a 97% R-squared, and each variable measured as highly significant, to 99.9% confidence. The model fit well on CO and CR pools across multiple coupons, and appears to continue to track well when predicting out of sample. One final caveat to the accuracy of the model is our exclusion of any measure of dollar roll specialness. Within our theoretical framework, dollar roll specialness should have a significant inverse relationship with loan balance pay-ups. That is, a highly special dollar roll should reduce or eliminate the carry advantage of slower prepayments in the specified pool collateral. However, calculating a historical measure of roll specialness is inherently subjective, and our simple approach did not return a significant or logical relationship to pay- up