Batch mode active learning framework and its application on valuing large variable annuity portfolios

In practice, the valuation of a large volume variable annuity contracts relies on Monte Carlo simulation which is quite computationally intensive. To build a more efficient valuation process, statistical models have been used within a data mining framework that consists of two subsequent stages: the data sampling stage to create a set of representative contracts, and the regression modeling stage to make predictions for the remaining contracts in the portfolio. In this article, we work with a new data mining framework based on active learning, in which we iteratively update the regression model efficiently by selecting the most informative representatives. Our metrics take into consideration both the ambiguity and the diversity of the prediction, which allow us to propose two methods that fit well in this active learning framework. Experimental results demonstrate the effectiveness of the proposed active learning approaches over the random sampling as well as the two-stage data mining framework.