Modelling financial stress during the COVID-19 pandemic: Prediction and deeper insights

Abstract

We model the evolutionary patterns of financial stress (FS) for the USA, other advanced economies (OAE), and emerging market (EM) regions during the COVID-19 pandemic. We propose an AI-driven framework to draw meaningful and actionable insights. A set of technical indicators and several allied macroeconomic features are chosen as explanatory variables filtered using the BorutaShap algorithm. We use Uniform Manifold Approximation and Projection (UMAP) based unsupervised feature processing to obtain a better representative structure. The set of engineered features is used to construct a Bidirectional Long Short-Term Memory Network (BLSTM) model to estimate the future values of FS across three regions. The outcome suggests that FS can be predicted during the extremely challenging COVID-19 pandemic, which is vital to evaluate the readiness toward digital transition of conventional financial services. The social media sentiment-based feature is significant for the US and OAE. The macroeconomic construct of WTI crude oil price is substantial for OAE and EM.