Trend-Error Decomposition for Self-Supervised Time Series Learning in Multivariate Forecasting Task

Self-Supervised Learning (SSL) has become a powerful paradigm in Artificial Intelligence, enabling the training of machine learning models using unlabeled data. However, in time series forecasting, SSL models are generally less effective than supervised models due to the complexity of temporal patterns, including trends, seasonality, and noise. To address this, we introduce TED4STL (Trend-Error Decomposition for Self-supervised Time-series Learning), a pipeline that decomposes each time series into two additive components, trend and error, and empirically test whether this decomposition improves the performance of SSL models. We adapt it to four SSL forecasting models and evaluate it on ten datasets. Experiments show that the decomposition consistently improves SSL forecasting accuracy, narrowing the gap with state-of-the-art supervised models and often surpassing them at short horizons.