Skip to main navigation Skip to search Skip to main content

Machine Learning for Early Detection of Credit Spread Regime Shifts in Emerging Markets

  • Marco Antonio Johannes Sarkady

    Student thesis: Master's Thesis

    Abstract

    Emerging market (EM) credit spreads widen abruptly under geopolitical, commodity,
    and macroeconomic shocks, raising bond-portfolio risk faster than a monthly investment
    committee can react. This thesis develops and evaluates a short-term early warning sys
    tem for the onset of spread-widening regimes across five regional JPMorgan Emerging
    Markets Bond Index (EMBI) sub-indices, covering Asia, Latin America, Eastern Eu
    rope, the Middle East, and Africa, using more than twenty years of daily market and
    macroeconomic data.
    Spread-widening onsets are identified through a hybrid rule-based and manually re
    viewed labeling procedure and predicted within a 21-day horizon. The dataset is severely
    imbalanced, with onsets at 5 to 10 percent prevalence and only 8 to 14 events per re
    gion in the validation window, so the study builds an evaluation framework around
    average-precision lift and event recall under a walk-forward protocol with causal alert
    thresholds. Logistic regression, random forest, and gradient-boosting classifiers are com
    pared against TiRex, a pretrained time-series foundation model, and daily scores are
    aggregated into a monthly committee signal of roughly one alert month per quarter per
    region for tactical asset allocation.
    Feature-engineered models reach average-precision lifts of 1.45 to 2.87 on the out-of-fold
    validation window and exceed naive spread- and VIX-momentum baselines, which stay
    at or below random. At the monthly committee layer, Asia, Eastern Europe, and Latin
    America retain strong ranking, with lifts of 2.60 to 3.68, while Africa and the Middle
    East fall to random. TiRex reaches a pooled lift of 1.50 but remains an exploratory ref
    erence, since its anomaly score was selected on overlapping data, which also exposes the
    evaluation challenges of pretrained forecasters in finance. These are validation-window
    estimates; the 2022 to 2025 holdout contains only two to five onsets per region, too few
    for firm out-of-sample confirmation. The contribution is a rare-event evaluation frame
    work and evidence that engineered EM spread features carry early-warning signal in
    three of five regions.
    Date of Award2026
    Original languageEnglish (American)
    Awarding Institution
    • FH OÖ
    SupervisorUlrich Bodenhofer (Supervisor)

    Studyprogram

    • Data Science and Engineering

    Cite this

    '