Insurance activities intrinsically deal with the management of risks. According to the mark-to-model approach to the assessment of the insurer’s debt position, the models used to forecast the probability structures involved in the computation of the fair value of the liabilities are central. We use artificial neural networks to the purpose of forecasting in this context. We focus on a portfolio of endowment insurance policies and on the ex ante estimation of the related mathematical reserve. We consider the Lee-Carter model for the forecasting of the random number of the policies in-force at each future policy anniversary, with either linear time-series model or autoregressive neural network models (ARX-NN models)for the time index. We find out that ARX-NN models allow to reduce the bias that linear time series models may imply. In particular, narrower prediction intervals around central forecasts translates into a much milder upward shift of the portfolio reserve in the case where systematic reduced mortality than expected should occur along the insurance contracts’ duration.