SBIR Phase I: Predictive and Computational Technologies for the Mortgage Industry

点击翻译按钮获取中文摘要

The broader impact of this Small Business Innovation Research (SBIR) Phase I project will result from providing mortgage market participants such as lenders, servicers, insurers, investors, rating agencies, government sponsored enterprises, and regulators with integrated deep learning systems that offer actionable predictions of borrower, portfolio, security, and market behavior of unprecedentedly high accuracy and low latency at scale. The systems will enable these organizations to identify valuable opportunities, reduce losses, and improve staff utilization while dramatically lowering compute costs in a $2 billion annual mortgage decision and risk analytics market. Wide adoption will boost the performance of the American housing-finance system, benefiting homeowners and the broader population through lowering borrowing costs, expanding access to credit, and reducing the risk of future financial crises. The Phase I project focuses on developing transformative computational algorithms that make comprehensive deep learning predictions available in real time, at a fraction of the cost of existing computational technologies. It yields new insights into how computational algorithms can significantly enhance the benefits of AI prediction systems. This Small Business Innovation Research (SBIR) Phase I project seeks to address the core technical challenge associated with the development of powerful deep learning systems for measuring risk and identifying opportunities in the mortgage industry. This challenge is the construction of novel and transformative asymptotic-approximation algorithms to run in real time, rather than the hours or days prior technology requires. Examples of such applications include measuring the risks of large pools of mortgages over long horizons and the risks of mortgage securities backed by such pools. The dramatic running time gains offered by these algorithms are the key to harnessing the unprecedented predictive accuracy of deep learning models of individual borrower behavior. The key objectives of the proposed project are to develop a large-pool asymptotic approximation approach that offers run-time guarantees for deep learning models, and to construct efficient numerical schemes for implementing the resulting algorithms in a cloud environment.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.