ABCD-Forecast:Augmentation and Bagging method for Confidential Data series Forecasting

Abstract

Financial time series prediction with machine learning is an important research topic both practically and academically. Financial time series are noisy, non-stationary, and may contain confidential information, which makes them more troublesome for researchers. To deal with these challenges, we propose a novel competition-based prediction method called Augmentation and Bagging method for Confidential Data series Forecasting (ABCD-Forecast). Our approach is inspired by the framework of data science competitions where multiple analysts submit their predictions and receive the feedback. ABCD-Forecast first distributes various de-noised versions of the data to virtual analysts, enabling the generation of diverse datasets without noise. Combining the predictions of these virtual analysts through a competition format allows us to obtain diverse and accurate models. Our method is applicable for different situations to handle non-stationary data. Furthermore, preprocessing and distributing the dataset through our method ensures data confidentiality, which is substantial in many actual situations. Empirical analysis using real-world data demonstrates the effectiveness of the proposed method in achieving good prediction accuracy.