An Algorithmic Crystal Ball: Forecasts-based on Machine Learning

IMF Working Papers, November 1, 2018

Source details

Canonical URL

An Algorithmic Crystal Ball: Forecasts-based on Machine Learning

Other formats

• Markdown version
• Structured JSON version
• Bundle manifest

Bibliographic details

• Authors: Jin-Kyu Jung, Manasa Patnam, Anna Ter-Martirosyan
• Published: November 1, 2018
• Series: IMF Working Papers
• DOI: https://doi.org/10.5089/9781484380635.001

Research overview

• Forecasting macroeconomic variables is key to developing a view on a country's economic outlook.
• Traditional forecasting models rely on fitting data to a pre-specified relationship between input and output variables, assuming a specific functional and stochastic process.
• The paper pursues a data-driven approach using machine learning algorithms that impose limited restrictions on the nature of the true relationship between input and output variables.
• Algorithms applied: Elastic Net, SuperLearner, and Recurring Neural Network.
• Empirical application: macro data for seven, broadly representative, advanced and emerging economies.

Main findings

• Machine learning algorithms can outperform traditional statistical models in forecasting macroeconomic variables.
• These algorithms offer a relevant addition to the field of economic forecasting by improving predictive performance relative to conventional approaches.
• The study specifically evaluates performance against benchmark WEO performance and WEO forecast comparisons (keywords indicate benchmarking to WEO).

Methodology and data notes

• Algorithms implemented: Elastic Net, SuperLearner, Recurring Neural Network.
• Data context: time series, training data, test data, macro variables including real GDP and GDP growth nowcasts.
• Comparative approaches referenced: decision tree algorithm, Random Forest algorithm, Random Forest algorithms to nowcast GDP growth.
• Emphasis on limited structural assumptions about the generation process of forecasts.

Implications and relevance

• Machine learning provides a complementary toolkit for macroeconomic forecasting tasks, particularly for nowcasting and forecasting real GDP.
• Results support exploring machine learning models alongside traditional econometric models for national economic outlook assessments.

Content in this bundle

• An Algorithmic Crystal Ball: Forecasts-based on Machine Learning, WP/18/230, November 2018
  • An Algorithmic Crystal Ball: Forecasts-based on Machine Learning, WP/18/230, November 2018 (Markdown version)
  • An Algorithmic Crystal Ball: Forecasts-based on Machine Learning, WP/18/230, November 2018 (PDF)