The main objective of this work is to define a statistical-learning model that combines data from multiple sources to predict cumulative production for a set of horizontal wells targeting Vaca Muerta formation. The set of predictive features includes geologic or reservoir information, geonavigation data and stimulation parameters for each well. Geologic information is represented by seismic attributes of different types extracted along well trajectories. A thorough exploratory analysis is conducted over the input data to understand the distribution and variability of proposed features. Outliers are identified and removed, and the dimensionality of the problem is reduced by evaluating collinearity effects and performing a principal component analysis. In this first stage of the work, we propose a multiparametric linear model through LASSO method (Least Absolute Shrinkage and Selection Operator), which allows for an optimum feature selection while training the model. Hence, the method contributes to further dimensionality reduction and to avoid overtraining effects, therefore balancing the bias-variance tradeoff. The linearity of the model helps interpreting the effect of each feature on the target variable, and its coefficients represent the importance of variables. We conclude that the trained statistical-learning model can explain a significant fraction of the well production data, while revealing synergy between geological properties and stimulation parameters. One of the advantages of using seismic data as features in the model is that it can be used to construct predictive maps of cumulative production. These maps could be interpreted as an indicator of productivity potential due to the quality of reservoir.