Unifying Asset Ranking and Portfolio Weighting through a Multi-Task Neural Network

This study presents a novel approach that integrates asset ranking and portfolio weighting within a single framework. Unlike traditional methods, which separate asset ranking from portfolio weighting, this research employs a multi-task neural network to concurrently learn asset rankings and optimize the number of assets for long and short positions. This innovation aims to better align the investment strategy with investor preferences right from the model prediction phase. To assess its effectiveness, the authors conduct experiments using historical weekly market data from China A-shares. The findings indicate that incorporating portfolio weighting into a multi-task learning framework significantly improves out-of-sample financial performance in contrast to benchmark methods that rely on heuristics or historical estimations.