Anomaly Detection of COVID-19 Impact on the US Stock Market using LSTM and Minimum Spanning Trees

Abstract

This paper investigates anomalous behaviors in the U.S. equity market during the COVID-19 pandemic by integrating Long Short-Term Memory (LSTM)–based anomaly detection with a network-theoretic visualization framework. We construct stock-specific LSTM models trained on pre-pandemic daily returns and identify anomalies in the test period by applying a threshold to the prediction-error distribution. This approach enables the model to highlight irregular return patterns that deviate from historically learned dynamics and are often associated with periods of heightened uncertainty or structural breaks. To enhance interpretability, the detected anomalies are mapped onto correlation-based Minimum Spanning Trees (MSTs) generated from normalized returns. By examining the MSTs across consecutive three-month windows from 2020 to 2021, we uncover clear temporal and topological transitions in market structure that correspond to the evolving stages of the pandemic. The earliest phase (T1) exhibits a dense concentration of anomalies and simple, pruning-like MST configurations, reflecting severe market disruption and rapid shock transmission. Subsequent phases display progressively more complex and stabilized network structures, indicating adaptation and partial recovery as market participants assimilated new information and adjusted to pandemic conditions. Our empirical findings, based on the Dow-30 constituents, demonstrate that combining LSTM-driven anomaly detection with MST visualization provides an intuitive yet rigorous framework for understanding crisis-induced market dynamics. This integrated approach offers decision-relevant insights into how systemic shocks propagate through financial networks and how market structure evolves under prolonged stress.