Leveraging augmentation techniques for tasks with unbalancedness within the financial domain: a two-level ensemble approach

Abstract Modern financial markets produce massive datasets that need to be analysed using new modelling techniques like those from (deep) Machine Learning and Artificial Intelligence. The common goal of these is forecast the behaviour market, which can translated into various classification tasks, such as, for instance, predicting likelihood companies’ bankruptcy or in fraud detection systems. However, it often case real-world data are unbalanced, meaning classes’ distribution not equally represented datasets. This gives main issue since any model trained according majority class mainly, leading inaccurate predictions. In this paper, we explore different augmentation deal with very unbalanced data. We consider a number publicly available datasets, then apply state-of-the-art strategies them, finally evaluate results several models on sampled performance approaches evaluated their accuracy, micro, macro F1 score, by analyzing precision recall over minority class. show consistent accurate improvement achieved when employed. obtained look promising indicate efficiency tasks. On basis results, present an approach focused tasks within domain takes dataset as input, identifies what kind technique use, applies ensemble all identified type input along methods tackle underlying classification.