A Neural Network Approach for the Estimation of Mortgage Prepayment Rates

The owner of a mortgage can decide to repay his loan before the contractual maturity, completely or in part, in most cases without suffering any penalty. This might happen for a number of reasons which can go even beyond market driven incentives - ie. a more favourable general level of interest rates; as a direct consequence, the client behavior can modify in a material way the future cashflows that the mortgage issuer would expect to receive according to a purely contractual view. Thus, the estimation of such behavioral effects is a key factor in order to guarantee an accurate representation of the future asset profile, which in turn is at the base of core activities of a bank such as asset and liability management, interest rate risk assessment and mitigation, as well as liquidity management.We present a behavioral model for prepayment risk based on a Deep Learning approach in which the rate of early repayment on the mortgage portfolio is estimated by means of a Neural Network. The Network is calibrated using the past history of the observed repayment events; the explanatory variables (ie. features) are selected via a Random Forest approach among both contractual elements and market factors. The model is applied to a sample portfolio of contracts, considering both floating-rate as well as fixed-rate mortgages. The resulting predictions are compared with the actual repayment rates observed in the past, proving that the Neural Network approach is suitable for this kind of behavioral modeling. Also, the model shows that the inclusion of the market rate levels as explanatory factors might be beneficial in estimating the repayment phenomenon also in the case of floating rate mortgages, and in particular when a fixed-rate mortgage that would be attainable for an equivalent loan become favourable, as in an environment of extremely low interest rates