In this study, we propose a novel framework for time series forecasting that combines the principles of diffusion models with traditional neural architectures to enhance predictive accuracy. Our method begins by introducing controlled Gaussian noise to an input time series, which is subsequently denoised through an iterative reverse diffusion process. This process yields a denoised version of the input sequence, which is then combined with the original sequence. The combined sequence is fed into different traditional architectures to predict future values. We argue that this hybrid approach offers more accurate forecasts compared to using the corresponding traditional neural architecture alone.