Transformer models have shown strong performance in the field of time series prediction, but they also have a significant “frequency bias” problem. Although the model learns high-energy, low-frequency signals, it still can easily ignore high-frequency signals that are critical to the prediction results. Some existing models usually solve this problem by splitting the spectrum into multiple bands, and performing local normalization and independent processing on each band. This strict band isolation strategies could eliminate bias, but they also prevent the model from capturing potential key dependencies between different bands, e.g., how long-term trends affect the amplitude of short-term fluctuations. To address this limitation, this paper proposes a cross-band fused FreqWave-Former, which introduces a two-layer encoder in a Transformer-based framework. First, debiased local representations are learned in isolated bands to inherit the debiasing advantages of existing methods. Then, a dedicated cross-band fusion attention module is designed to model relationships between these bands in a controlled and unbiased manner, which can effectively prevent the model from ignoring lower amplitude features. Extensive experiments conducted on 16 benchmark datasets demonstrate that our model significantly outperforms other state-of-the-art baseline models, validating our hypothesis that prediction accuracy can be improved by modeling cross-frequency dynamics without reintroducing the original frequency bias.