Classic Theta model decomposes the original data series into two separate lines, which are extrapolated separately and the forecasts are combined with equal weights. The current study explores a neural network approach to Theta model, in terms of optimizing the combination weights of the two components in the final forecast. The performance of the proposed method (Theta AI) is compared against the original method for the two subsets of the NN3 forecasting competition, which primary objective was the evaluation of methods using neural networks or artificial intelligence for time series forecasting. Results indicate that the new approach is very promising towards the generalization of the Theta model.