CrisisPredictor: An Explainable AI Early-Warning Monitor for Financial Stress

Financial crises are costly, and the signals that warn of them come in different units from different sources: credit spreads, market volatility, jobs data and the news. This paper describes CrisisPredictor, an open and explainable monitor that reads 14 such signals, asks a language model to score each one from −10 to +10 against written scoring rules retrieved from a small knowledge base, and then combines the scores with fixed weights written in code into one 0–100 stress score. The language model never sets the final number. We test the score on 300 historical dates drawn at random from stress and calm periods chosen in advance (150 of each), using real historical market and economic data and real New York Times headlines for the two news factors. At the live alarm line of 55 the score was right on 219 of 300 dates (73%; 95% range 68–77%), caught 89 of 150 stress dates and rang falsely on 20 of 150 calm dates, with an area under the ROC curve (AUC) of 0.86 (95% range 0.82–0.90). The news factors added almost nothing: replacing AI-generated news summaries with real headlines changed the AUC by only 0.008, and VIX used alone separated stress from calm better than the full score (AUC 0.93). Tests of advance warning were negative: over 226 month-ends the score predicted a further 10% market fall within six months with an AUC of only 0.57, and a trend model chosen on 2006–2015 fell to 0.57 and 0.43 on unseen years. We therefore present CrisisPredictor as a transparent stress detector and risk monitor, and we report its advantages (explainability, fixed and auditable weights, low cost, full reproducibility) separately from its limits, which we state in full. A live version of the monitor is available at https://145-241-213-220.sslip.io.