Trigger-amplifier-resilience dynamics of short-horizon systemic risk: An interpretable machine-learning early warning system for Indonesia

This research develops an interpretable early warning system for short-horizon systemic risk using the Trigger-Amplifier-Resilience (TAR) framework. Rather than treating crises as static outcomes, the framework considers them as evolving processes, where shocks in financial markets act as initial triggers, are subsequently transmitted through financial linkages, and are ultimately shaped by underlying macro-financial conditions. Using monthly Indonesian data from 2015 to 2024, the analysis applies a Random Forest model and uses SHAP decomposition to unpack the drivers behind predicted risk. Crisis episodes are identified based on market-based stress thresholds, and model performance is evaluated through an expanding-window out-of-sample approach. The results suggest that the machine-learning model improves crisis detection relative to a linear benchmark, pointing to the relevance of nonlinear dynamics. More importantly, the decomposition reveals a consistent sequencing pattern: market signals tend to move first, amplification mechanisms follow with a short delay, and resilience factors become more relevant as stress unfolds. Taken together, these findings indicate that combining prediction with interpretation can provide more useful real-time signals, allowing policymakers to respond more precisely to different phases of systemic stress.