Alleviating Distribution Shift in Time Series Forecasting with an Invertible Neural Network Transformation

The distribution of time series data changes over time, posing challenges for accurate time series forecasting. One common approach to tackle the issue of distribution shift involves transforming the data into a latent space where the impact of the shift is minimized. However, existing methods heavily depend on experienced distribution assumptions and lack guidance on the latent space, leading to sub-optimal performance enhancements. To tackle the above challenges, we propose a new transformation technique to explicitly mitigate the distribution shift between historical and forecast data without any distribution assumptions. Specifically, an Invertible Neural Network Transformation INNT is designed to convert data into a smooth latent space. The INNT is constructed to be bidirectional and reversible by a temporal slicing mechanism, thereby preserving all information from the original data. Moreover, the transformation process is guided by a pretraining strategy that aims at reducing distribution divergence within the latent space. Additionally, the proposed method is model-agnostic, allowing for seamless integration into various existing forecasting models. Extensive experiments are conducted to validate the accuracy and generalization of the proposed framework.