Amid the increasing complexity and intelligence of financial markets, there is an escalating imperative for investors to anticipate multi-dimensional indicators of return and risk. Conventional forecasting methodologies in finance predominantly target single objectives, thereby failing to capture the comprehensive dynamics of asset performance and limiting their utility in multi-objective investment decision-making. To address this limitation, we propose MPFI-Net (Multi-prediction model in financial investment-Network), a multi-task financial indicator forecasting framework grounded in the Transformer architecture. Initially, financial time series are encoded through an embedding layer combined with positional encoding, enabling the Transformer encoder to extract global temporal features. Subsequently, a multi-task output module is designed to concurrently predict three pivotal indicators—return, volatility, and maximum drawdown—while a dynamic loss-weighting strategy is incorporated to enhance synergistic learning across tasks. Empirical evaluations demonstrate that MPFI-Net substantially surpasses traditional single-task models and established neural network approaches across both U.S. and China market datasets, achieving superior accuracy and robustness in terms of mean squared error (MSE) and mean absolute error (MAE) metrics. This approach provides an effective and robust modeling paradigm for multi-objective financial forecasting, with promising potential for practical application in investment decision-making.
Shao et al. (Mon,) studied this question.