Total: 1
In this paper, we present a novel feature extraction procedure to predict interval-valued time series by combing transfer learning and imaging approaches. Initially, we represent interval-valued time series using a bivariate point-valued time series, which serves as a representative form. We first transform each time series into images by employing various imaging approaches such as recurrence plot, gramian angular summation/difference field, and Markov transition field, and construct an image dataset by treating each imaging method's output as a separate class. Based on this dataset, we train several candidates for a feature extraction network (FEN), specifically ResNet with varying layers. Then we choose the penultimate layer of the FEN to extract the most relevant features from the transformed images. We integrate the extracted features into conventional predictive models to formulate the corresponding prediction models. To formulate prediction, we integrate the extracted features into a regular prediction model. The proposed methods are evaluated based on the S\&P 500 index and three data-generating processes (DGPs), and the experimental results demonstrate a notable improvement in prediction performance compared to existing methods.