AbstractIn recent years, fused lasso models are becoming popular in several fields, such as computer vision, classification and finance. In portfolio selection, they can be used to penalize active positions and portfolio turnover. Despite efficient algorithms and software for solving non-smooth optimization problems have been developed, the amount of regularization to apply is a critical issue, especially if we have to achieve a financial aim. We propose a data-driven approach for learning the regularization parameters in a fused lasso formulation of the multi-period portfolio selection problem, able to realize a given financial target. We design a neural network architecture based on recurrent networks for learning the functional dependence between the regularization parameters and the input data. In particular, the Long Short-Term Memory networks are considered for their ability to process sequential data, such as the time series of the asset returns. Numerical experiments performed on market data show the effectiveness of our approach.