Enhance Volatility of Denormalized Predictions in Time Series Forecasting

In time series forecasting, mainstream models commonly normalize multivariate representations of the same timestamp for stable results. However, directly outputting denormalized results may lead to over-smooth predictions. These denormalized predictions suffer from the bias of amplitude scale and occasionally deviate far from actual ground truth. To alleviate this issue, we propose a novel time series forecasting model, Friformer, which compensates for potential volatility after denormalization by forecasting a seasonal part in the frequency domain. We leverage the real-valued fast Fourier transform to capture frequency domain information and employ complex-valued linear layers to predict future local fluctuation sequences. Additionally, we introduce a variable token-based attention mechanism to enhance the prediction of future trends. Through comprehensive experiments on seven benchmark datasets, our proposed method reduces the forecasting error by about 11% compared with state-of-the-art baselines.