Moving beyond conventional mean-targeted forecasts, I predict the entire distribution of bond risk premia to capture extreme, crisis-driven tail movements. Augmenting standard predictors with an extracted upper-tail macroeconomic factor—representing a latent "macro-panic"—achieves full predictive dominance over baseline models across the quantile spectrum. Methodologically, a Quantile Random Forest (QRF) efficiently captures non-linear crisis thresholds using mean-selected predictors, offering a computationally tractable, high-performing alternative to Quantile LASSO. The resulting forecasts exhibit pronounced regime-dependent cyclicality, becoming heavily countercyclical during recessions. Crucially, this approach reveals a term-structure duality: short-term tail risks are driven by acute macro-liquidity panics ("dash for cash"), whereas long-term risks are anchored by structural safe-haven demand ("flight to safety"). Finally, I document a decay in macroeconomic predictability along the yield curve: while the QRF dominates at shorter maturities, naive empirical quantiles become highly competitive at the long end.