Forecasting of time series implies time and effort in order to achieve accuracy. In this work is proposed a new strategy of times series forecasting based on the extraction components of low and of high frequency. The strategy imply four stages, embedding, decomposition, estimation and recomposition. In the first stage, is used the Hankel matrix to embed the original time series. In the second stage is applied the Singular Value Decomposition (SVD) technique, with SVD are extracted the components of low and high frequency. In the third stage is implemented an Autoregressive Neural Network (ANN) based on Particle Swarm Optimization (PSO), the ANN makes the estimation of the components. The recomposition is the final stage and here is obtained the forecasted value, here is computed with the single addition of the estimated components obtained in the third stage. The evaluation of this proposal is developed with two time series of traffic accidents occurred in Concepci´on Chile, from year 2000 to 2012, the data sampling period is weekly. The results obtained are compared with the values given by the conventional forecasting process, showing the high accuracy and superiority of this proposal, the gain in MAPE is 498:3% and the gain in R2 is of 53:2% for the time series injured people, and with the time series injured people was obtained a gain of similar proportion. Keywords: Autoregressive Neural Network Particle Swarm Optimization Singular Value Decomposition