VaR and CVaR-Based Stress Testing Using Deep Learning for Liquidity Risk Forecasting and Banking Stability Assessment

Sazzadul Islam; Rebeka Sultana

doi:10.63125/291phs66

Authors

Sazzadul Islam Master of Science in Big Data Analytics, Bay Atlantic University, Washington, USA Author
Rebeka Sultana Master of Arts in Information Technology Management, Webster University, TX, USA Author

DOI:

https://doi.org/10.63125/291phs66

Keywords:

Liquidity risk, Deep learning (LSTM), Value at Risk (VaR), Conditional Value at Risk (CVaR/Expected Shortfall), Banking stability (Z-score)

Abstract

Liquidity risk remains a core driver of banking fragility because funding shocks and market illiquidity can force value destructive asset sales and destabilize cash flow obligations. This study addresses the problem that conventional liquidity ratios and linear models can understate tail risk, weakening early warning and stress testing for stability oversight. The purpose was to validate an integrated pipeline combining deep learning liquidity risk forecasting with VaR and CVaR based stress testing to strengthen banking stability assessment. In a quantitative cross-sectional, case-based design, four enterprise case banks (B = 4) were analyzed using 192 aligned bank observations and a survey of 220 liquidity risk professionals, producing 212 usable responses (96.4%). Key variables included liquidity buffer ratio (liquid assets/total assets), wholesale funding reliance, loan-to-deposit ratio, a Likert 1-5 governance and stress testing maturity index, VaR95 and VaR99, CVaR95 and CVaR99, and a stability proxy (Z-score). The analysis plan covered descriptive profiling, forecasting comparison (LSTM versus linear regression), tiered scenario stress testing (baseline, mild, adverse, severe), and correlation and regression models linking stressed tail risk and governance to stability. Governance measurement was reliable (Cronbach’s alpha = 0.81-0.89) and scenario realism was rated high (M = 4.08, SD = 0.58). The LSTM improved prediction accuracy over the benchmark (RMSE = 0.042 vs 0.061; MAE = 0.031 vs 0.047) and directional accuracy (67.5% vs 56.2%). Under severe stress, VaR0.95 rose from 1.90 to 2.83 (+48.9%) while CVaR0.95 rose from 2.66 to 4.39 (+65.0%); VaR0.99 rose from 3.12 to 4.71 (+51.0%) while CVaR0.99 rose from 4.48 to 7.96 (+77.7%), showing tail thickening captured by CVaR. Stressed CVaR0.99 correlated negatively with stability (r = -0.58, p < .001) and remained significant in regression (β = -0.46, p < .001; R² = .47). Governance predicted lower stressed CVaR (β = -0.29, p = .002) and moderated the CVaR to stability link (interaction β = +0.18, p = .010). Implications are that CVaR stress testing supports funding structure and liquidity buffer decisions, while governance maturity reduces liquidity vulnerability.