Beyond Static Prepayment Curves: Machine Learning and Simulation for Prepayment and Liquidity Risk in Government-Guaranteed Loan Pools

Abstract

However‚ for the guaranteed portion of a loan‚ a borrower default is an early return of principal to the investor‚ making static prepayment curves structurally unfit to price SBA 7(a) guaranteed portions and Ginnie Mae pools of FHA and VA loans․ Replacing them with machine learning is necessary‚ but not sufficient․ It shows that loan level prediction contributes to valuation under an accelerated simulation structure‚ a governance layer for validators to contest․ Our three contributions are: (i) a taxonomy of prepayment mechanisms for guaranteed pools‚ (ii) a comparison mechanism for benchmark static‚ structural‚ gradient boosted‚ deep neural network and survival models via five criteria‚ and (iii) a reference architecture coupling prepayment prediction with servicer liquidity stress․ The timing of the repurchase‚ and the funding capacity of the servicer‚ should be modeled jointly‚ since their effects are related in a way that makes separate modeling misleading․