A Metadata-Augmented Technique for Time Series Forecasting and Transfer Learning

Enhancing the performance of forecasting and transfer learning is a continual challenge for time series data. Although metadata, which provides additional context and infor-mation about the primary data, has the potential to enhance time series forecasting substantially, this aspect is still underexplored. In this work, we propose a metadata-augmented technique for time series forecasting. The technique involves augmenting the original time series with metadata obtained by computing the first difference of consecutive data points in the series once the time series forecasting has been formulated as a supervised learning problem. Experimental results show that the use of our metadata-augmented technique outperforms the use of the original data alone for both time series forecasting and transfer learning. This demonstrates that transforming the data using our technique can significantly improve performance in various time series applications.