Generalized Meta Framework for Forecasting

Forecasting tabular time series data has become a challenging task as the time series data has its own unique patterns, and therefore identifying the most suitable modeling approach for a given dataset requires additional investigations and expert knowledge. In this study, we propose a novel meta framework that utilizes an ensemble approach, combining the models with a high level of performance and efficiency for a given dataset with the aim of proposing a more generalized framework for time series forecasting. In this proposed approach, we use different models from a large pool of candidate models that have the ability to capture unique timeseries characteristics available in various time series data sets. This approach allows us to get a more robust and generalized framework. By using a variety of forecasting models, including statistics-based prediction models, machine learning-based prediction models, deep learning-based prediction models, and generative modeling, our novel approach ensures broad applicability across various datasets from different application domains. In the proposed approach, meta features that help describe the structure, complexity, and time series patterns available in a given dataset are used to determine the optimal ensemble from a large pool of candidate models for the final prediction process. In the past literature, it was noted that the generative models, such as Variational Autoencoders (V AEs) and Generative Adversarial Networks (GANs), demonstrate a high level of performance in high-dimensional data contexts; however, their usability in the time series data context is limited due to their incapability of capturing temporal dependencies in the time series data. In order to address this limitation, we use a hybrid architecture to incorporate generative models into our pool of candidate models. Then a stacking approach was used to integrate the predictions of the ensemble's member models as it conditionally weighted the predicated values. Through this approach, our focus was on proposing a more robust, generalized forecasting approach across diverse time series datasets with various time series features more effectively. We used various synthetic and real-world tabular time series data sets to show the generalizability and wide applicability of our proposed framework in different application domains.