Accurate Survival Prediction

点击翻译按钮获取中文摘要

Many tasks involves estimating the time until an event: eg, doctors often need to predict the time until a patient**will die (or until she has a relapse, or until she recovers); retail companies want to predict the time until a**customer will leave ("customer churn"), or until an employee will resign. The field of survival analysis has**produced a great many tools for analyzing this data, but most generate a score that can be used to rank patients**rather than an explicit prediction of survival time.**In previous research, we've developed a method called Patient-Specific Survival Prediction (PSSP), which we**use to predict the probability of survival at any desired time of interest, as well as compute the expected**survival time. Further, we have developed metrics to evaluate the accuracy of survival analysis methods (a**challenge of survival data is that frequently we know only a lower-bound of an instance's survival time, but not**the exact survival time, making evaluation difficult).**The Royal Bank of Canada is excited to apply PSSP (and other survival analysis methods) to the area of credit**risk. Specifically, RBC is interested in applying the results of this project to its installment loan and auto**finance products in the effort of predicting at what time an individual is likely to default on a loan. Predicting**when a client will default or pre-pay the loan are paramount in calculating the viability of underwriting the**loan.**The goals of the proposed project are to explore how to optimize accuracy of predicted survival time, and to**explore the use of ensemble models to improve upon individual survival analysis methods, within the specific**domain of credit risk.