The Causal Impact of COVID-19 on G20 Stock Markets Using LSTM Counterfactual Forecasting

Abstract

The COVID-19 pandemic has significant implications to the financial stability and stock market value. Researchers have examined the pandemic effects using econometric models, this approach struggles to capture the nonlinear dynamics, structural breaks, and extreme volatility associated with pandemic periods. This paper utilized deep learning techniques to model complex time-series under unprecedented shocks. This paper proposed a Long Short-Term Memory (LSTM) neural network for time-series prediction to predict and calculate causal inference for G20 countries from 1 January 2018 to 30 December 2020. The analytical framework incorporates pandemic-related variables: daily confirmed cases (DCC), daily death cases (DDC), and the stringency index (SINDEX), macroeconomic indicators such as exchange rate (ER), interest rate (BOND), volatility index (VIX), gold price (GOLD), and oil price (OIL). The experiments also evaluate the important features for each model by using SHapley Additive exPlanations (SHAP). The results show that the LSTM model achieved a high predictive accuracy for all G20 countries. The causal analyses reveal heterogeneous pandemic effects across G20 markets, with significant negative impacts observed in France, the United Kingdom, and Indonesia, while most developed markets demonstrated stronger resilience. The proposed LSTM model achieved consistently strong performance across most countries, with MAPE values ranging from 2.07% to 5.84%.