Unleashing the Unpredictable: Generating Context-Driven Synthetic Black Swans

Black swans are rare high-impact crisis events that disrupt societies and financial markets, e.g., the 2008 global financial crisis. Due to their rarity, it is challenging to predict them with certainty using traditional statistical methods. Research on modeling black swans explored the use of deep learning techniques, such as LSTMs, Autoencoders, and self-supervised methods. However, data scarcity remains a challenge for both supervised and unsupervised methods. We present SwanSynthetiX, an approach for generating context-driven synthetic black swan events in open domains that closely resemble real-world extreme event data. It combines Extreme Value Theory (EVT) with conditional GANs (cGANs) and addresses data scarcity through EVT-based Monte Carlo-sampling. The approach uses scenarios to capture unique extreme event circumstances, enabling cGANs to generate context-driven black swans with distinct characteristics. Experiments demonstrate SwanSynthetiX outperforming recent approaches in synthetic time-series generation, empowering signal detection in crisis management for early identification of real-world black swan events.