A composite distribution is a two-component spliced distribution, which generally equals a lighter-tailed probability density function (pdf) up to a threshold and a heavy-tailed pdf thereafter. Such distributions are used to model heavy-tailed actuarial loss data characterized by a large number of claims with small size and few claims with large size. In this study, we consider the Lognormal-Pareto composite regression model obtained by introducing covariates in the Lognormal-Pareto composite distribution. The statistical inference for such a model is very challenging due to the unknown threshold where the composite distribution changes its shape. Therefore, we approach the prediction problem by means of neural networks with the purpose to obtain a prediction interval for the expected value of this regression model; this expected value is the basis of the insurance premium. We discuss the results of this technique on some generated data sets.