Predicting sales using cannibalization effects is a current major challenge. Several approaches deal with the estimation of cannibalization of new product launches or promotion effects but how to account for it when predicting future sales is still missing. For this reason, we aim to fill this gap by proposing a new framework based on time series causality as a method to identify potential candidates of causality. To attain such an objective, we use two state-of-art gradient boosting based algorithms, namely Extream gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LGBM ), as well as two Multi-step forecasting strategies. We show that the cannibalization approach together with Recursive forecasting provides more accurate forecasts respect to established benchmark models.