This paper presents a novel approach to stock ranking forecasting and portfolio construction through a regime-dependent multi-task learning framework. By using auxiliary tasks to aid the main list-wise risk-adjusted returns ranking task, we enhance the model’s accuracy and stability. These auxiliary tasks provide additional gradients and valuable insights, such as predicting future stock movements and volatility changes. Moreover, our model adapts to different market regimes—bullish, bearish, turmoil, or tranquil—by emphasizing relevant factors and model structure for each condition. By considering the first moment and the second moment conditional on both market-level regimes and stock-level tasks, we highlight that our model provides an elegant and suitable method for risk-adjust portfolio construction. Empirical results using Chinese stock market data demonstrate significant improvements in the accuracy and robustness of portfolio performance, offering a robust tool for investors.