Data reconstruction of two actuarial metrics by staking machine learning models

Objective: A large part of Brazilian’s health care is financed by health insurance plans, which readjustments have been questioned in the courts. The data from court cases tends to not be readily available. Therefore, in order to reconstruct the data, we developed a metric using Deep Learning techniques to obtain data estimations. Method: After analyzing the data obtained from the Regulatory Agency, we trained three different supervised learning algorithms aiming to obtain information through an optimization problem. We used the Augmented Lagrangian method aiming to include the constraints into the cost function and Simulated Annealing to minimize it. Results: Consistent as expected, the stacking performance outperformed the base learners. Conclusions: With the results obtained it was possible to obtain the retroactive average cost per claim and frequency information, fetched from the "health plan's past".