Methodological Insights into Structural Causal Modelling and Uncertainty-Aware Forecasting for Economic Indicators

Federico CeruttiarXiv 2025 · DOI · arXiv

Introduction

Economic indicators such as GDP and unemployment are components of a complex, non-linear system, making them notoriously difficult to predict. Traditional forecasting methods often suffer from two primary limitations: the "Black Box" problem, where models provide predictions without explaining the underlying drivers, and the "Uncertainty Gap," where point estimates fail to account for the range of possible outcomes or extreme shocks. This paper addresses these challenges by integrating Structural Causal Modelling (SCM) with uncertainty-aware probabilistic forecasting.

Methodology

The proposed framework consists of a two-stage pipeline:

  1. Causal Discovery: The Latent PCMCI (LPCMCI) framework is employed to identify causal links within time series data. To capture non-linear dependencies common in economic systems, the Gaussian Process Distance Correlation (GPDC) test is utilized. This approach allows for the discovery of complex relationships that standard linear tests might overlook.
  2. Zero-Shot Forecasting: The framework leverages Chronos, a foundation model for time series. Chronos treats numerical values as discrete tokens, similar to language models, and was pre-trained on a vast corpus of diverse time series data. This enables "zero-shot" forecasting—predicting future values for unseen datasets without additional task-specific fine-tuning.

Data and Preprocessing

The study analyzes four key U.S. macroeconomic indicators from 1970 to 2021: Gross Domestic Product (GDP), Economic Growth (quarterly % change in GDP), Inflation (monthly CPI change), and the Unemployment Rate. Data was sourced from the OECD and the U.S. Bureau of Economic Analysis (BEA), with standard scaling applied to ensure numerical stability during modeling.

Causal Insights

The causal analysis provides a structural justification for the forecasting phase. A key finding is the strong autoregressive structure of Unemployment, indicating that its future values are heavily influenced by its own historical patterns. In contrast, Inflation appeared relatively isolated within this specific variable set, suggesting its primary drivers may be latent factors not captured in the four-variable model, such as global supply chain shocks or commodity prices.

Causal graph illustrating the temporal dependencies between macroeconomic indicators.

Forecasting Performance and Uncertainty

The integration of Chronos allows for probabilistic forecasting, providing a 90% prediction interval that serves as a diagnostic tool for anomalies. When observed data falls outside this interval, it signals a structural shift or an unprecedented event, such as the 1982 recession or the COVID-19 pandemic.

The zero-shot approach demonstrated high accuracy, with an average predictive accuracy of 0.929 at a one-quarter horizon, maintaining a robust 0.762 even four steps ahead. The 90% prediction intervals covered approximately 81% of observations, a reasonable calibration given the extreme economic shocks present in the historical data.

Distribution of absolute prediction errors across four forecast horizons.

Conclusion

By combining causal discovery with large-scale pre-trained time series models, this research demonstrates a path toward more interpretable and robust economic forecasting. While the current univariate approach shows significant promise, future work will focus on multivariate probabilistic forecasting to better capture how causal shocks propagate through the economy in real-time.

Summary updated

BibTeX
@misc{cerutti2025methodological,
  author = {Federico Cerutti},
  title = {{Methodological Insights into Structural Causal Modelling and Uncertainty-Aware Forecasting for Economic Indicators}},
  year = {2025},
  eprint = {2509.07036},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.07036},
}
Figure