The reliability of sales forecasting is critical for an industrial decision support system dedicated to raw material retailers. However, it turned difficult to train and maintain a custom model dedicated to each of the numerous references. For every reference, it would be needed to select the most accurate algorithm together with its relevant features, then, to exhaustively test every relevant combination of parameters. This was the reason why we explored an approach based on auto parametrization of well-known predictive models, while adding specific seasonal features. From our experiments, the Dynamic Harmonic Regression (DHR) based on ARMA stood out as being the most effective model for popular products: it reached a fair accuracy while requiring a reasonable cost to train. However, when it came to more volatile products, a simple prediction like the average sales per week over a year often performed the best. Thus, YearlyMean saved computational resources that could then be used to exhaustively train DHR or LSTM models on some company key products, leading to a potential improvement of their forecasts. Then, one details the implementation of a smart computing machine learning process based on predictive scenarios that seek for a trade-off between the consumed resources and the predictive performances.