Machine Learning for Economics Research: When What and How? (2304.00086v2)

Abstract: This article provides a curated review of selected papers published in prominent economics journals that use ML tools for research and policy analysis. The review focuses on three key questions: (1) when ML is used in economics, (2) what ML models are commonly preferred, and (3) how they are used for economic applications. The review highlights that ML is particularly used to process nontraditional and unstructured data, capture strong nonlinearity, and improve prediction accuracy. Deep learning models are suitable for nontraditional data, whereas ensemble learning models are preferred for traditional datasets. While traditional econometric models may suffice for analyzing low-complexity data, the increasing complexity of economic data due to rapid digitalization and the growing literature suggests that ML is becoming an essential addition to the econometrician's toolbox.

• The paper presents a systematic review of when, what, and how to apply machine learning in economics, focusing on methodological rigor and practical implementation.
• It compares deep learning, ensemble, and causal ML models, emphasizing their strengths in handling complex, nontraditional datasets for enhanced predictive accuracy.
• The paper highlights strategies such as transfer learning and tailored model use while addressing challenges like computational demands and issues with interpretability.

Machine Learning for Economics Research: When, What, and How?

The paper by Ajit Desai, published under the auspices of the Bank of Canada, presents a meticulous review of the utilization of ML within the field of economics. Specifically, the work addresses three pivotal questions: when is ML applied in economics, what models are preferred, and how are these models employed in economic research and analysis.

Contextual Relevance of Machine Learning in Economics

The integration of ML into economic research is underscored by the burgeoning digitization of the economy, which has led to an exponential increase in both the volume and complexity of data. This evolution necessitates analytical methods that can efficiently handle large-scale and nontraditional datasets. Desai highlights that while traditional econometric models remain effective for low complexity data, ML is gaining traction due to its ability to manage complex, nontraditional data and improve predictive accuracy. The paper shows a discernible trend, underscored by an increase in ML models being cited and applied in leading economics journals, indicating ML's growing utility in the field.

Preferred Machine Learning Models and Their Relevance

The selection of ML models in economic research largely depends on the characteristics of the data and the specific application at hand:

1. Deep Learning: These models are particularly advantageous for handling nontraditional data such as text, images, and audio. For example, latent Dirichlet allocation (LDA) is frequently used for text analysis in topic modeling. More advanced deep learning architectures like Transformers are being utilized for their capability to handle large datasets and complex structures in natural language processing tasks.
2. Ensemble Learning: Preferred for traditional datasets, ensemble learning models such as random forests and boosting methods are employed to improve robustness and predictive accuracy. These models are especially effective in scenarios involving multicollinearity and nonlinearity in data, providing a comprehensive approach to understanding complex economic phenomena.
3. Causal ML Models: When the objective is causal inference, causal ML models provide the methodological tools necessary to analyze large datasets without heavily relying on predefined structures, which can lead to uncovering causal relationships hidden within data.

Application and Implementation Strategies

The paper outlines pragmatic approaches in employing ML in economic contexts. These include using pre-trained models and transfer learning, especially beneficial when dealing with deep learning models that require extensive datasets and computational resources. Furthermore, tailoring ensemble learning models to specific tasks rather than employing them generically enhances their efficacy. For example, in macroeconomic forecasting, adapting models to suit the dataset's temporal and structural variations yields better predictive power.

Practical and Theoretical Implications

The use of ML in economics, as highlighted by Desai, brings numerous advantages, particularly for large and nontraditional datasets. Theoretical implications include the enhanced ability to model nonlinearity and observe complex interactions within data, which may remain obscured in traditional econometric models. Practically, these advances offer economists tools to better inform policy analysis, providing more accurate forecasts and improved understanding of economic systems.

Challenges and Prospects

Despite the promising integration of ML into economics, Desai underscores several limitations. The data-intensive nature of ML, the necessity for substantial computational resources, and challenges in model interpretability present significant hurdles. Furthermore, standard statistical properties for ML models have yet to be fully established. Future research may focus on enhancing model interpretability and developing methodologies to confer standard errors and asymptotic properties, thus broadening the scope of ML's applicability in economics.

In conclusion, the paper offers a comprehensive overview of the substantial contributions and limitations of ML within the field of economic research, providing a foundation for future exploration into optimizing ML applications in this field. The discussion encourages ongoing dialogue and research into how ML can continue to evolve and enhance the toolkit available for economists, addressing both complex and traditional datasets.