Paper: arXiv 2505.16136

Authors: Yuke Zhang

Abstract

This study introduces an interpretable machine learning (ML) framework to extract macroeconomic alpha from global news sentiment. We process the Global Database of Events, Language, and Tone (GDELT) Project’s worldwide news feed using FinBERT – a Bidirectional Encoder Representations from Transformers (BERT) based model pretrained on finance-specific language – to construct daily sentiment indices incorporating mean tone, dispersion, and event impact. These indices drive an XGBoost classifier, benchmarked against logistic regression, to predict next-day returns for EUR/USD, USD/JPY, and 10-year U.S. Treasury futures (ZN). Rigorous out-of-sample (OOS) backtesting (5-fold expanding-window cross-validation, OOS period: c. 2017-April 2025) demonstrates exceptional, cost-adjusted performance for the XGBoost strategy: Sharpe ratios achieve 5.87 (EUR/USD), 4.65 (USD/JPY), and 4.65 (Treasuries), with respective compound annual growth rates (CAGRs) exceeding 50% in Foreign Exchange (FX) and 22% in bonds. Shapley Additive Explanations (SHAP) affirm that sentiment dispersion and article impact are key predictive features. Our findings establish that integrating domain-specific Natural Language Processing (NLP) with interpretable ML offers a potent and explainable source of macro alpha.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: reported Sharpe ratio of 5.87 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 8.0)

Why this score: The paper employs advanced machine learning models (XGBoost, FinBERT) and SHAP for interpretability, but focuses more on application and backtesting rather than deep theoretical derivations. It demonstrates high empirical rigor with a decade-long OOS backtest (2015-2025), realistic transaction costs, and published code on GitHub, though the Sharpe ratios presented seem exceptionally high and warrant scrutiny.

Research Flowchart

  flowchart TD
  A["Research Goal: Extract Macro Alpha<br>via Interpretable ML on News Sentiment"] --> B["Data Source: GDELT Global News Feed"]
  B --> C["NLP: FinBERT Sentiment Processing"]
  C --> D["Feature Engineering:<br>Sentiment, Dispersion, Impact"]
  D --> E["ML Model: XGBoost Classifier"]
  E --> F["Out-of-Sample Backtesting<br>(5-Fold Expanding Window)"]
  F --> G{"Key Outcomes"}
  G --> H["High Performance:<br>Sharpe ~5.87 (FX), ~4.65 (Bonds)"]
  G --> I["SHAP Analysis:<br>Dispersion & Impact are Top Features"]
  G --> J["Conclusion:<br>Interpretable ML provides<br>Robust Macro Alpha"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes an interpretable machine learning framework that uses FinBERT sentiment indices and an XGBoost classifier to generate macroeconomic alpha from global news.

Source sentence
This study introduces an interpretable machine learning (ML) framework to extract macroeconomic alpha from global news sentiment.

Assumes 6

  • Trading signals are generated from predicted probabilities with a 0.5 threshold and positions are taken at the next day's open. execution
    Source sentence
    a long position (+1 unit) is initiated for day t + 1 if ˆpt > 0.5, and a short position (-1 unit) is initiated if ˆpt ≤0.5.
  • Transaction costs of 0.02% per round‑trip for FX and 0.05% for Treasury futures are deducted from returns. execution
    Source sentence
    Transaction costs are deducted from these returns: 0.02% of the traded value per round‑trip for FX pairs and 0.05% for ZN Treasury futures
  • Only daily headline sentiment is used, ignoring intraday news dynamics. data
    Source sentence
    We perform sentiment analysis exclusively on the extracted headlines to maximize robustness and computational efficiency.
  • Out‑of‑sample evaluation uses an expanding‑window scheme from early 2017 to April 2025. statistical
    Source sentence
    Our backtesting framework uses data from January 2015 through April 2025 with an expanding window methodology for OOS evaluation.
Show 2 more
  • Statistical significance is assessed via a block bootstrap with 1,000 resamples and 20‑day blocks. statistical
    Source sentence
    we employ a block bootstrap procedure (e.g., 1,000 resamples with a block size of 20 trading days, approximating one month of trading data to preserve autocorrelation)
  • The binary target is 1 if the next‑day log return is positive, otherwise 0. model
    Source sentence
    The binary prediction target yt for day t is defined as 1 if the next day’s return rt+1 > 0

Data

Backtest sample: 2017 – 2025-04 · EUR/USD · fx · daily

  • GDELT news headlines · Global Database of Events, Language, and Tone (GDELT) Projec text-news public
    2015‑2025 · global news articles
    Source sentence
    Our pipeline ingests daily news from the Global Database of Events, Language, and Tone (GDELT) Project, applies FinBERT to score each article’s sentiment, and aggregates these scores into a suite of features
  • FX spot rates (EUR/USD, USD/JPY) · Yahoo Finance market-prices public
    2015‑2025 · EUR/USD and USD/JPY currency pairs
    Source sentence
    FX spot rates (EUR/USD, USD/JPY) are sourced from Yahoo Finance (tickers EURUSD=X, USDJPY=X)
  • 10‑year U.S. Treasury futures (ZN) · Yahoo Finance market-prices public
    2015‑2025 · continuous front‑month 10‑year Treasury futures
    Source sentence
    A continuous front‑month 10‑year Treasury futures series (ZN) is also sourced from Yahoo Finance (ticker ZN=F)

Recognised sources: GDELT Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

0.02% per round‑trip for FX pairs and 0.05% per round‑trip for ZN Treasury futures are subtracted from daily strategy returns.

Source sentence
Transaction costs are deducted from these returns: 0.02% of the traded value per round‑trip for FX pairs and 0.05% for ZN Treasury futures

Finds 5

Reported performance (XGBoost news sentiment trading strategy): Sharpe 5.87 · annual return 55.4% · total return 8989.3% · max drawdown 15.6%

  • XGBoost achieves out‑of‑sample Sharpe ratios of 5.87 (EUR/USD), 4.65 (USD/JPY) and 4.65 (10‑yr Treasury) with CAGR above 50% for FX and 22% for bonds. performance
    Source sentence
    Sharpe ratios achieve 5.87 (EUR/USD), 4.65 (USD/JPY), and 4.65 (Treasuries), with respective compound annual growth rates (CAGRs) exceeding 50% in Foreign Exchange (FX) and 22% in bonds.
  • XGBoost substantially outperforms logistic regression, which yields Sharpe ratios below 1 and negative CAGR for Treasuries. performance
    Source sentence
    The XGBoost models demonstrate exceptionally strong and consistent performance across all targeted assets, significantly outperforming the logistic regression baseline.
  • SHAP analysis identifies sentiment dispersion and article impact as the most important predictive features. predictive-accuracy
    Source sentence
    SHAP analysis confirmed that the models rely on economically meaningful, news‑derived features as primary drivers
  • The strategy remains robust across subperiods and under realistic transaction costs. statistical
    Source sentence
    results persisting across subperiods and cost assumptions
  • A daily‑frequency model may miss intraday news effects, noted as a limitation. negative
    Source sentence
    A primary limitation of this study is its reliance on daily frequency data.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 30 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.