This paper introduces FreDF, a frequency-enhanced direct forecast method for time series forecasting. Time series modeling faces challenges due to autocorrelation in both historical data and future sequences. While current methods focus on historical data autocorrelation, future label correlations are often overlooked. FreDF addresses this by learning to forecast in the frequency domain, reducing estimation bias. Experiments show FreDF outperforms existing methods and is compatible with various forecast models. The key contributions include identifying label autocorrelation as a critical challenge, proposing FreDF to mitigate this issue, and validating its effectiveness through extensive experiments. FreDF aligns forecast and label sequences in the frequency domain, effectively reducing label autocorrelation bias. It maintains the advantages of the Direct Forecast (DF) paradigm, including sample efficiency and simplicity. FreDF is model-agnostic and compatible with various forecasting models, enhancing performance across diverse time series forecasting scenarios. The method is evaluated on multiple datasets, demonstrating its effectiveness in improving forecast accuracy and reducing bias. The paper also discusses the theoretical basis of FreDF, including the impact of label autocorrelation on learning objectives and the benefits of frequency domain forecasting. The results show that FreDF significantly improves forecast performance, particularly in long-term forecasting tasks. The method is flexible and can be applied to various forecasting models, making it a valuable tool for time series forecasting.This paper introduces FreDF, a frequency-enhanced direct forecast method for time series forecasting. Time series modeling faces challenges due to autocorrelation in both historical data and future sequences. While current methods focus on historical data autocorrelation, future label correlations are often overlooked. FreDF addresses this by learning to forecast in the frequency domain, reducing estimation bias. Experiments show FreDF outperforms existing methods and is compatible with various forecast models. The key contributions include identifying label autocorrelation as a critical challenge, proposing FreDF to mitigate this issue, and validating its effectiveness through extensive experiments. FreDF aligns forecast and label sequences in the frequency domain, effectively reducing label autocorrelation bias. It maintains the advantages of the Direct Forecast (DF) paradigm, including sample efficiency and simplicity. FreDF is model-agnostic and compatible with various forecasting models, enhancing performance across diverse time series forecasting scenarios. The method is evaluated on multiple datasets, demonstrating its effectiveness in improving forecast accuracy and reducing bias. The paper also discusses the theoretical basis of FreDF, including the impact of label autocorrelation on learning objectives and the benefits of frequency domain forecasting. The results show that FreDF significantly improves forecast performance, particularly in long-term forecasting tasks. The method is flexible and can be applied to various forecasting models, making it a valuable tool for time series forecasting.