A Boxplot Metadata Configuration Impact on Time Series Forecasting and Transfer Learning

The potential impact of metadata on enhancing the accuracy and reliability of time series forecasting is increasingly being recognized. Including metadata in forecasting models provides a more comprehensive understanding of the underlying data, leading to improved model performance and more accurate predictions. This paper aims to investigate the impact of metadata configuration on time series forecasting, with a focus on exploring the impact of using metadata derived from the boxplot configuration with comparison to raw data for time series forecasting and transfer learning performance. The boxplot information configured as metadata, as suggested in this work, entails converting the lag features and/or the horizon values into an 8-value data chunk each, consisting of the five boxplot values of the series, plus the first, middle, and last values of the series. For the majority of the lag feature and horizon value combinations taken into consideration in this experiment, the boxplot metadata consistently outperforms the raw data in both time series forecasting accuracy and transfer learning performance. However, further analysis and experimentation are needed to gain a more complete understanding of all the potential benefits and applicability of boxplot-derived metadata in time series forecasting.