An Interpretable Generative Framework for Anomaly Detection in High-Dimensional Financial Time Series

A new research paper introduces ReGEN-TAD, an interpretable generative AI framework designed to tackle the notoriously difficult problem of spotting anomalies in high-dimensional financial time series. Authored by Waldyn G. Martinez, the model addresses the dual challenges of complex temporal dependencies and evolving cross-sectional relationships between assets. It integrates modern machine learning with established econometric diagnostics within a refined convolutional-transformer architecture. The core innovation is its dual-task approach, performing both joint forecasting and reconstruction of financial data, which allows it to build a robust understanding of normal market behavior.

ReGEN-TAD generates four distinct, complementary signals to flag anomalies: predictive inconsistency (when forecasts fail), reconstruction degradation (poor data recreation), latent distortion (shifts in hidden model states), and volatility shifts. These signals are robustly calibrated and aggregated into a single, unified anomaly score, crucially without needing any pre-labeled anomalous data for training. Experiments on synthetic and real financial panels show the framework improves robustness against structured market deviations. Most importantly for finance professionals, it enables economically coherent, factor-level attribution, meaning it can explain which underlying economic drivers (like interest rates or sector performance) are responsible for the detected instability, moving beyond a simple alert to provide actionable diagnostic insight.