In a classical time series forecasting task, the first standard decision when modeling involves the adoption of statistical methods or other pure machine learning models, including three based algorithms or deep learning techniques. The choice is strongly related to the problem we are carrying out but in general: statistical techniques are adequate when we face an autoregressive problem when the future is related only to the past; while machine learning models are suitable for more complex situations when it's also possible to combine variegated data sources. In this post, I try to combine the ability of the statistical method to learn from experience with the generalization of deep learning techniques. Our task is a multivariate time series forecasting problem, so we use the multivariate extension of ARIMA, known as VAR, and a simple LSTM structure. We don't produce an ensemble model; we use the ability of VAR to filter and study history and provide benefit to our neural network in predicting the future.