CDFM

Cluster of Dynamic Factor Models for macroeconomic forecasting

The Cluster of Dynamic Factor Models (CDFM) bridges large-scale structural macroeconomic models and flexible small-scale factor models. It consists of a series of sector-specific dynamic factor models sequentially linked within a cluster structure. These links are established through Granger-causality tests using classical and neural-network-based techniques. Estimated using a Kalman filter, the CDFM generates disaggregated GDP forecasts across the production, expenditure and income sides of the National Accounts using mixed-frequency monthly and quarterly indicators. The framework accommodates expert judgment and scenario analysis while maintaining interpretability and empirical robustness.

The Cluster of Dynamic Factor Models (CDFM) bridges the gap between large-scale structural macroeconomic models and flexible, small-scale factor models. The CDFM is constructed as a series of small-scale dynamic factor models, each tailored to a specific economic sector or variable, and sequentially linked within a cluster structure.

The connections between these models are established using Granger-causality tests, drawing on both classical and modern (including nonlinear neural network-based) techniques to identify robust directional relationships among variables. This architecture enables the CDFM to exploit the cross-correlation structure of time series data, while maintaining the interpretability and consistency characteristic of structural models.

The CDFM generates disaggregated GDP forecasts across the quarterly production, income, and expenditure sides of the System of National Accounts (SNA) using a wide range of monthly and quarterly indicators. Each small-scale factor model within the cluster is estimated using the Kalman filter, which offers several practical advantages: it can handle missing observations and mixed-frequency data and allows for conditional forecasts.

The CDFM not only produces consistent and accurate forecasts across a wide set of macroeconomic variables but also allows for the incorporation of expert judgment and scenario analysis. Overall, the CDFM provides a flexible, interpretable, and empirically robust framework for large-scale macroeconomic forecasting, combining the strengths of both structural and data-driven approaches.

Further information available for download

Presentation "Macroeconometric forecasting using a cluster of dynamic factor models" (PDF)