Learning A Causation-driven Retrieval Model for Effective Time Series Augmentation and Forecasting

For time series forecasting models, retrieving internal or external time series as auxiliary information may help enhance the models' predictive power and robustness when forecasting target time series. However, retrieving time series based on their observed parts often leads to unreliable predictions because of their natural randomness: the time series similar in the current time window may behave differently in the future. To overcome this issue, in this study, we propose a simple but effective causation-driven retrieval method, called CaRTS, for time series augmentation and forecasting. In particular, we propose to “retrieve causes from effect”, matching time series based on the similarities between their prediction parts. To overcome the inaccessibility of the prediction parts, we train a dual-tower retrieval model to encode the observed parts of matched time series, making the similarities between their latent representations inherit the similarities between their prediction parts. By leveraging this retrieval model, we can enhance the prediction model's performance by augmenting each target time series with retrieved data likely to exhibit similar future behaviors. To provide concrete evidence of its effectiveness, we rigorously evaluate the CaRTS method on real-world datasets. Experiments show that CaRTS is a promising approach to retrieving time series with high temporal correlations, leading to higher prediction accuracy, improving the model performance and robustness significantly.