From Dot-com to AI: drawdown onsets prediction in AI technology stock market

Abstract

This paper introduces the Iterative Market Understanding and Generation Inference (IMUGI) framework to anticipate the onset of major drawdowns in AI technology stocks at their peak, rather than identifying them once the decline has already unfolded and the hedging window has closed. The study's purpose is to establish whether drawdown-onset prediction with actionable lead time is feasible by combining multiple crisis mechanisms into a unified online probabilistic architecture, rather than relying on the single-mechanism specifications common in the existing literature. IMUGI is a three-layer probabilistic architecture comprising: (1) a Second-Order Hidden Markov Model (SOHMM) with Student-t emissions and Missing-Not-At-Random (MNAR) handling for latent regime inference; (2) a Quadratic Hawkes process for endogenous feedback and extreme-return clustering; and (3) a regularised logistic meta-model that fuses both signals into a daily drawdown-onset probability. Daily price data for 27 NASDAQ-listed AI and technology firms spanning 2003–2025 were retrieved via the yfinance API. The framework is evaluated on a strictly held-out test set of 999 trading days (November 2021 -- October 2025), with predictions generated incrementally to prevent look-ahead bias. The framework attained an AUC-ROC of 0.9055 and a Brier score of 0.1175 on the held-out test set. All three drawdown onsets in the test period are detected: two are flagged with advance warning of five and one trading days, respectively, and one is identified within two trading days of onset. The recency of the AI technology sector as a distinct investable field limits the number of observable drawdown onsets; the three events in the test window should be read as statistically constrained but consistent out-of-sample evidence rather than a comprehensive test of generalisation.