Collateral Disruption Reconsidered: A Deep Machine Learning Framework for Modeling Third-Party Sanctions Spillovers, Trade Diversion Dynamics, and De-Risking Behavior in Pakistan's Cross-Border Banking Compliance Regime

Abstract

Purpose. Pakistan is not the target of any comprehensive sanctions programme, yet over two decades its banks, payment channels and trade routes have absorbed the costs of measures aimed at others, chiefly Iran and Russia, while three spells on the Financial Action Task Force (FATF) grey list and two US penalties on Pakistani banks raised compliance costs at home. This paper asks how large these third-party, or "collateral", disruptions are, through which channels they travel, and whether deep learning adds anything that conventional econometrics misses.

Design/methodology/approach. We build an annual dataset for 2005–2024 that combines World Bank macro-banking indicators, UN Comtrade bilateral trade with six partners, and an author-coded sanctions-compliance pressure index. A two-stage framework first identifies spillovers with bootstrap ARDL bounds tests, nonlinear ARDL, dynamic OLS, Jordà local projections, a Minnesota-prior Bayesian VAR with Diebold–Yilmaz connectedness, Hansen threshold regression, double/debiased machine learning and Toda–Yamamoto causality. It then trains LSTM, GRU, two-layer attention-LSTM and deep multilayer-perceptron networks, a deep ensemble and a denoising autoencoder, evaluates them out of sample with Diebold–Mariano tests and a model confidence set, and explains them with exact group-Shapley values.

Findings. Compliance pressure is associated with a persistent contraction of private credit: the dynamic-OLS long-run semi-elasticity is −25.3 percentage points of GDP per index unit, the bootstrap bounds test finds a levels relationship at the 10% level, and pressure Granger-causes credit but not the reverse. A one-standard-deviation pressure shock cuts reserves by about 0.7 months of imports on impact and lowers credit by up to 0.45 points of GDP after two years. Iran sanctions intensity shifts recorded trade away from Iran and towards UAE intermediation. Remittances follow a threshold regime: above a pressure index of about 0.46 they become sensitive to currency depreciation, consistent with leakage to informal channels. Deep models reduce forecast error by up to 22% relative to a prevailing-mean benchmark, but no gain is statistically significant in eleven annual hold-out years. Their value lies in attribution: sanctions channels account for 32–46% of the deep ensemble's explanatory mass, with Iran sanctions dominating trade diversion and US enforcement and the Russia regime dominating the 2022–2024 remittance shortfall.

Originality/value. The paper is the first to integrate a full small-sample econometric identification strategy with deep sequence models and exact Shapley attribution for a third-country sanctions spillover problem, and it reports, rather than conceals, the limits of deep learning on short macro samples.