Can speculative bubbles be identified (and timed) before they crash? This project builds a catalog of 69 U.S. equity sector episodes (1994–2024): 27 bubbles that crashed more than 40% and 42 "near-bubble" controls that rallied 50%+ but held, scored on 23 metrics spanning equity volatility, SEC filing language, bond credit-implied volatility, and a novel text-based credit signal. Using survival analysis and 10,000 leverage-stratified holdout splits with strictly pre-peak data, the full model identifies bubble sectors at 0.844 AUC (permutation p = 0.033), while equity volatility alone is statistically indistinguishable from noise.

Timing, however, fails completely: apparent month-level predictability is an artifact of the episode clock, and within fixed windows every feature performs at chance; crash triggers are genuinely exogenous. The result is a detection tool, not a prediction engine: it flags sustained leveraged fragility that can persist for years before a catalyst arrives.