Paper: arXiv 2506.22763

Authors: Fiona Xiao Jingyi, Lili Liu

Abstract

Forecasting central bank policy decisions remains a persistent challenge for investors, financial institutions, and policymakers due to the wide-reaching impact of monetary actions. In particular, anticipating shifts in the U.S. federal funds rate is vital for risk management and trading strategies. Traditional methods relying only on structured macroeconomic indicators often fall short in capturing the forward-looking cues embedded in central bank communications. This study examines whether predictive accuracy can be enhanced by integrating structured data with unstructured textual signals from Federal Reserve communications. We adopt a multi-modal framework, comparing traditional machine learning models, transformer-based language models, and deep learning architectures in both unimodal and hybrid settings. Our results show that hybrid models consistently outperform unimodal baselines. The best performance is achieved by combining TF-IDF features of FOMC texts with economic indicators in an XGBoost classifier, reaching a test AUC of 0.83. FinBERT-based sentiment features marginally improve ranking but perform worse in classification, especially under class imbalance. SHAP analysis reveals that sparse, interpretable features align more closely with policy-relevant signals. These findings underscore the importance of integrating textual and structured signals transparently. For monetary policy forecasting, simpler hybrid models can offer both accuracy and interpretability, delivering actionable insights for researchers and decision-makers.

Complexity vs Empirical Score

  • Math Complexity: 5.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced ML architectures (XGBoost, transformers, SHAP) and statistical evaluation, but the core math is standard for quantitative finance; the empirical rigor is high due to a detailed dataset, specific performance metrics (AUC 0.83), and reproducible methodology.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Predict Fed Policy Moves"] --> B{"Methodology"}
  B --> C["Data Acquisition<br>FOMC Texts & Macro Indicators"]
  C --> D["Feature Engineering<br>TF-IDF, FinBERT, Economic Data"]
  D --> E["Model Training<br>Unimodal vs. Hybrid Models"]
  E --> F{"Evaluation & Analysis"}
  F --> G["Findings & Outcomes"]
  subgraph F ["Computational Processes"]
      F1["Classification<br>XGBoost, Deep Learning"]
      F2["Ranking & AUC Analysis"]
      F3["SHAP Interpretability"]
  end
  G --> G1["Hybrid Models Outperform"]
  G --> G2["TF-IDF + XGBoost<br>Best AUC: 0.83"]
  G --> G3["Simplicity + Interpretability<br>Actionable Insights"]

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 a hybrid model that combines TF‑IDF features of Federal Reserve communications with macroeconomic indicators using an XGBoost classifier.

Source sentence
The best performance is achieved by combining TF-IDF features of FOMC texts with economic indicators in an XGBoost classifier, reaching a test AUC of 0.83.

Assumes 2

  • Class imbalance is addressed by synthetic oversampling (SMOTE). data
    Source sentence
    The class imbalance problem is further addressed by integrating SMOTE, which synthetically augments minority class instances to promote equitable learning.
  • Sentiment probabilities from FinBERT are assumed to capture relevant textual signals despite class imbalance. model
    Source sentence
    FinBERT-based sentiment features marginally improve ranking but perform worse in classification, especially under class imbalance.

Data

  • Structured macroeconomic indicators · Federal Reserve Economic Data (FRED) macro public
    Source sentence
    ... All data are retrieved from the Federal Reserve Economic Data (FRED) database, maintained by the Federal Reserve Bank of St. Louis.
  • Federal Reserve communications texts · Board of Governors of the Federal Reserve System text-news public
    January 2011 to January 2025
    Source sentence
    ... These documents span from January 2011 to January 2025 and cover five key types of communication.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • Hybrid TF‑IDF + economic features with XGBoost achieve a test AUC of 0.83, outperforming unimodal baselines. predictive-accuracy
    Source sentence
    The best performance is achieved by combining TF-IDF features of FOMC texts with economic indicators in an XGBoost classifier, reaching a test AUC of 0.83.
  • Gradient Boosting model (XGBoost) attains Train AUC 0.9139, Test AUC 0.8116, and Test Accuracy 52.3% after SMOTE. performance
    Source sentence
    Our tuned Gradient Boosting model achieves excellent in-sample performance (Train AUC: 0.9139; Accuracy: 75.7%) and respectable generalization on the test set (Test AUC: 0.8116; Accuracy: 52.3%).
  • FinBERT‑based sentiment features provide marginal ranking improvements but lower classification accuracy, especially with class imbalance. predictive-accuracy
    Source sentence
    FinBERT-based sentiment features marginally improve ranking but perform worse in classification, especially under class imbalance.
  • Standalone text‑only models achieve limited performance (e.g., FinBERT AUC 0.6869, accuracy 0.6690). performance
    Source sentence
    FinBERT (fine-tuned) 0.6869 0.6690

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 31 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.

Cited References (5)

  1. Generalized autoregressive conditional heteroskedasticity (1986)
  2. Estimating the support of a high-dimensional distribution (2001)
  3. Isolation Forest (2008)
  4. Long short term memory networks for anomaly detection in time series (2016)
  5. MAD-GAN: Multivariate anomaly detection for time series data with generative adversarial networks (2019)

0 of 5 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.