Dataset Card for Dataset MacroEconHorizon

Dataset Details

This dataset provides a comprehensive collection of macroeconomic time series indicators designed specifically for benchmarking forecasting models on economic prediction tasks.

Dataset Description

MacroEconHorizon is a curated collection of macroeconomic time series spanning multiple countries from 1970 to 2023. The dataset is specifically designed to support nowcasting, forecasting, and scenario analysis tasks for machine learning researchers and economic policy makers. It contains distinct time series across 63 countries, capturing a comprehensive range of indicators including GDP, inflation, unemployment rates, commodity prices, financial market variables, and monetary policy metrics.

• Curated by: [Bank for International Settlements]
• License: [CC BY-NC-ND 4.0]

Uses

Direct Use

The MacroEconHorizon dataset is intended for:

• Benchmarking time series foundation models on macroeconomic forecasting tasks
• Evaluating conditional forecasting capabilities where models predict under different policy scenarios
• Comparative analysis of forecasting performance across different economic regimes
• Testing models' ability to handle mixed-frequency data common in economic statistics
• Research on transfer learning approaches in economic forecasting

The dataset supports both unconditional forecasting (where future covariate values are unknown) and conditional scenario analysis (where specific paths for variables are imposed over segments of the forecast horizon).

Out-of-Scope Use

This dataset is not designed for:

• High-frequency trading or financial asset pricing
• Individual-level economic decision making
• Definitive policy guidance without expert interpretation
• Commercial use without permission

Dataset Structure

The dataset contains time series at daily, monthly, and quarterly frequencies, standardized into a unified format through our preprocessing pipeline. Each time series is identified by a unique code and includes the following metadata:

• Code: Unique identifier for each variable
• Description: Explanatory label for the economic indicator
• Frequency: Observation frequency (D=Daily, M=Monthly, Q=Quarterly)
• Transformation: Applied transformation type:
  • value: No transformation (raw values) - for already stationary series
  • value_delta: First difference (Δx^i_t = x^i_t - x^i_t-1) - for non-stationary series
  • value_log: Log (log(x^i_t)) - for series with positive values where proportional changes are relevant
  • valude_delta_log: Log difference (Δlog(x^i_t) = log(x^i_t) - log(x^i_t-1)) - for series with exponential growth

Dataset Creation

Curation Rationale

The dataset was created to address the need for a standardized benchmark for evaluating time series foundation models on macroeconomic forecasting tasks. Existing time series benchmarks often lack:

-Realistic publication lags that reflect real-world economic data availability -Mixed-frequency variables that create challenges for forecasting models -Designated test periods spanning different economic regimes -Support for scenario analysis as commonly used in policy settings

MacroEconHorizon bridges these gaps by providing a carefully curated dataset that reflects the practical challenges faced by economic forecasters.

Source Data

Data Collection and Processing

Raw data were collected from official sources and stored by Bank for International Settlements (BIS). The preprocessing pipeline includes:

• Transformation selection based on statistical properties of each series to achieve stationarity
• Temporal alignment via daily forward-fill operations to handle mixed frequencies
• Publication lag simulation by shifting time indices to reflect realistic data availability

Who are the source data producers?

The raw data comes from central banks, and international financial institutions that produce and publish macroeconomic statistics. These include the Bank for International Settlements (BIS), national central banks, and government statistical offices.

Personal and Sensitive Information

The public dataset contains only aggregate economic indicators with no personal or sensitive information. Variables that may be subject to confidentiality agreements or proprietary restrictions have been excluded from the public release and are marked accordingly in our documentation.

Bias, Risks, and Limitations

• Historical coverage: While the dataset spans from 1970 to 2023, coverage is not uniform across all variables and countries, with some series having shorter histories.
• Revision effects: Economic data is frequently revised after initial publication; the dataset represents the most recent vintage available and does not capture revision dynamics.

Recommendations

• Consider the appropriate transformation for each variable based on its statistical properties.
• Account for publication lags when designing realistic forecasting experiments.
• Be aware of the designated test periods when comparing model performance.
• Recognize that forecasting performance on historical data may not necessarily translate to future predictive accuracy.