Unveiling the Impact of Macroeconomic Policies: A Double Machine Learning Approach to Analyzing Interest Rate Effects on Financial Markets

Anoop Kumar, Suresh Dodda, Navin Kamuni, Rajeev Kumar AroraarXiv 2024 · DOI · arXiv

This research, presented at the 2024 IEEE International Conference on Artificial Intelligence for Internet of Things (AIIoT), applies advanced causal inference techniques to a fundamental question in finance: How significantly do Federal Reserve interest rate hikes impact investment returns?

TL;DR

Researchers utilized a Double Machine Learning (DML) framework to isolate the causal impact of US Federal Reserve interest rate changes on fund returns. The study reveals a stark contrast: while a 1% increase in interest rates causes a significant 11.97% drop in actively managed fund returns, the impact on passively managed funds is inconsistent and less predictable.

What Problem Does This Paper Address?

In financial modeling, distinguishing between correlation and causation is a major challenge. Traditional econometric tools often struggle with the "noise" and high-dimensional nature of modern financial data. While it is well-known that interest rates and market returns are related, proving that a specific policy caused a certain percentage of a fund's decline is difficult when numerous macroeconomic factors—such as GDP, inflation, and unemployment—are shifting simultaneously.

Correlation matrices for passive and active fund datasets, illustrating the complex interdependencies between macroeconomic variables that the DML framework must navigate.

The authors bridge the gap between predictive Machine Learning (which excels at pattern recognition) and Causal Inference (which identifies underlying drivers). They specifically investigate the "treatment effect" of interest rate hikes on two distinct groups: actively managed funds and passively managed funds.

Methods

The core of the research is the Double Machine Learning (DML) framework. DML isolates the true causal effect by "partialling out" the influence of confounding variables through two stages:

  1. The Y-Task: Predicting the outcome (fund returns) using control variables.
  2. The D-Task: Predicting the treatment (interest rate changes) using those same control variables.

By subtracting these predictions from the actual values, the model isolates the "residual" relationship, representing the isolated causal effect.

Key Technical Approaches: * Model Comparison: The study compared standard Linear Regression against Gradient Boosting (XGBoost) within the DML framework. * Data Preprocessing: To address non-stationary data, the researchers used first-differencing and the Augmented Dickey-Fuller (ADF) test. * Lag Analysis: Using Vector Autoregressive (VAR) models, they determined a seven-month lag was optimal for capturing the delayed effects of monetary policy. * Fixed Effects: Means-encoding was employed to handle the high cardinality of nearly 7,000 unique fund tickers without over-expanding the feature space.

Datasets

The study spans a comprehensive timeframe from January 1986 to December 2021, capturing multiple market cycles.

  • Fund Data: Sourced via Bloomberg, the initial pool of 14,816 funds was narrowed down to a final dataset of 1,948 passive funds and 5,000 active funds.
  • Macroeconomic Data: Sourced from FRED (Federal Reserve Economic Data), including M2 money supply, FRS interest rates, unemployment, CPI inflation, and GDP growth.
The systematic filtering process used to refine the initial fund universe into a robust dataset for analysis.

Key Insights and Ideas

  • Superiority of Non-Linear Models: Gradient Boosting significantly outperformed Linear Regression in predicting returns (the "Y-Task"), achieving an $R^2$ of 0.73 for active funds compared to just 0.13 for the linear model. This underscores that the relationship between macro-policy and markets is deeply non-linear.
  • Active vs. Passive Sensitivity: A major finding is the divergence in results. Active funds showed a highly significant, negative causal response to rate hikes. Conversely, passive funds yielded non-significant results in the Gradient Boosting DML model ($p = 0.301$), suggesting that the "passive" label may encompass too broad a range of strategies to exhibit a uniform causal response.
  • Systematic Policy Decisions: The models were exceptionally accurate at predicting interest rate changes (achieving $R^2$ values of 0.99). This suggests that Federal Reserve decisions are highly systematic and closely tied to the macroeconomic variables included in the study.

Key Results

  • Causal Impact: For actively managed funds, a 1% increase in interest rates causes an 11.97% decrease in returns.
  • Model Accuracy: Gradient Boosting reached an $R^2$ of 0.81 for passive funds and 0.73 for active funds in predicting returns, validating it as a robust learner for the DML framework.
  • Statistical Significance: The causal effect for active funds was supported by a very low p-value ($2e-6$), providing strong empirical evidence for macroeconomic theories regarding monetary policy sensitivity.

Limitations and Future Work

  • Data Constraints: Bloomberg terminal limits restricted the breadth of the active fund dataset to 5,000 funds per pull.
  • The "Passive" Mystery: The inconsistent results for passive funds suggest that future research should categorize them further (e.g., index funds vs. sector-specific ETFs) to pinpoint specific causal links.
  • Geographic Scope: The study is limited to the U.S. market. Validating this DML approach in international markets is a logical next step to strengthen the framework’s generalizability.
  • Computational Intensity: DML, particularly with Gradient Boosting and cross-validation, is computationally expensive, which limited the extent of parameter tuning.

Summary updated

BibTeX
@misc{kumar2024unveiling,
  author = {Anoop Kumar and Suresh Dodda and Navin Kamuni and Rajeev Kumar Arora},
  title = {{Unveiling the Impact of Macroeconomic Policies: A Double Machine Learning Approach to Analyzing Interest Rate Effects on Financial Markets}},
  year = {2024},
  eprint = {2404.07225},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.07225},
}
Figure