A THREE-STEP MACHINE LEARNING APPROACH TO PREDICT MARKET BUBBLES WITH FINANCIAL NEWS

Battle Count: 85

The paper provides a practical early warning framework for bubble detection that could inform risk management and position sizing decisions. The multilabel classification (bubble up vs. bubble down) is particularly relevant for directional trading strategies. However, the paper focuses on prediction rather than trading strategy implementation, does not include transaction costs or portfolio construction, and the sentiment features add minimal value. The macroeconomic indicators (GDP, CPI, BOP) are the dominant predictors, which are typically available with a lag. The framework is more suited for systemic risk monitoring than high-frequency trading.

The three-step pipeline requires: (1) implementing the PSY/GSADF test with proper window sizes and critical values, (2) deploying finbert-lc for sentiment scoring on large-scale news data, and (3) training ensemble models with hyperparameter grid search. The data preprocessing involves interpolation from monthly/quarterly to biweekly frequencies. The multilabel classification logic adds complexity. However, all components use well-established libraries (statsmodels for PSY, HuggingFace for finbert-lc, scikit-learn/XGBoost for ensemble methods). The main complexity lies in data collection and the threshold-dependent labeling scheme.

The paper describes the methodology in detail including PSY procedure, finbert-lc model, and ensemble methods. Data sources are identified (Bloomberg, Reuters, NYT, eodhd API). However, no code repository is provided. The PSY test parameters, interpolation methods, and specific hyperparameters for grid search are mentioned but not fully specified. The finbert-lc model is referenced from a prior publication by the same author.