While model-free forecasting via complicated machine learning techniques is often used for large portfolio prediction in financial sector, it loses the interpretability of the prediction procedure, and affects consumer protection and make the supervision challenge for prudential regulators. When it comes with the determining the price of financial product, most of countries regulate financial institutions’ valuation models to ensure that it is reasonable, understandable, and objective. As a result, a model-based approach is a much sensible choice. In this talk, we first deal with the least square Monte Carlo (LSMC) method, which combine Monte Carlo simulation and the ordinary least-square regression, for the model-based forecasting problem of insurance portfolio with large number of policyholders. Then, extending the LSMC method, we provide an efficient deep-learning algorithm of the model-based forecasting under the general statistical model assumption.