Transformer-Based Forecasting of Digital Transaction Volatility under Monetary Policy Shocks: Shock-Conditioned Attention Models with Regime- and Horizon-Wise Evidence

Digital payment ecosystems are increasingly exposed to abrupt volatility in transaction flows, especially during periods of monetary policy repricing and heightened financial uncertainty. This study proposes a shock-conditioned Transformer to forecast digital transaction volatility using a real-time, walk-forward evaluation design that preserves temporal causality and announcement timing. Empirically, the shock-conditioned model improves forecast accuracy across horizons, reducing aggregate RMSE from 0.0271 to 0.0250 (+7.7% relative improvement) and improving MASE from 0.91 to 0.84. Gains concentrate in policy-relevant short horizons: RMSE (h=1–3) declines from 0.0210 to 0.0190 (+9.5%). Under high shock intensity (top 10% st), RMSE improves from 0.0345 to 0.0304 (+11.9%) and tail-focused error on high-volatility days (top decile of yt) decreases by 13.6%. Regime stratification shows stronger benefits during tightening: RMSE improves by +9.3% in tightening versus +5.4% in easing. Ablation tests confirm mechanism validity: removing shock inputs eliminates most gains and increases short-horizon RMSE by 6.8–8.9% relative to the full model, while dropping shock–stress interaction terms disproportionately degrades performance during stress regimes. Diagnostic attention statistics show systematic concentration around shock windows, with mean attention mass rising from 0.24 in non-event periods to 0.37 within ±2 days of announcements. Overall, results indicate that policy shocks act as regime triggers for transaction volatility and that explicitly conditioning Transformers on shock innovations delivers robust, horizon-consistent improvements suitable for early warning and operational risk monitoring in digital payment systems.