Forecasting Financial Spillover Under Endogenous Trading Halts

Financial distress spreads through a network of firms, but the worst-affected firms often halt trading entirely, which censors exactly the data needed to measure their effect on others---a missing-not-at-random mechanism that standard stochastic dynamic methods do not handle. We present the halt-aware Temporal Edge Gaussian Process (HA-TE-GP), combining a bounded GP transition with sender-row particle filtering and a halt likelihood. Across 200 matched nonlinear-DGP replications, GP-transition methods substantially improve MSE and probabilistic scores over recursive boosting, linear, and persistence baselines, while the halt-aware variants yield no practically meaningful improvement over transition-only TE-GP.