Abstract
We adapt graphical causal structure learning methods to apply to nonstationary time series data, specifically to processes that exhibit stochastic trends. We modify the likelihood component of the Bayesian Information Criterion (BIC) score using a Vector Error-Correction Model (VECM) to ensure consistency in the presence of integrated and cointegrated variables. Our approach, integrated into the SVAR-GFCI algorithm, allows for the discovery of causal relationships even when latent confounders are present.
The Challenge of Nonstationarity in Causal Discovery
Most algorithms for learning causal graphs from time series data assume stationarity—the property that the statistical characteristics of the data do not change over time. However, real-world macroeconomic data, such as GDP, interest rates, or inflation, often exhibit nonstationarity. These variables frequently behave like "random walks," where their values drift over time without returning to a fixed mean.
A common practice in econometrics is to take the first difference of the data (e.g., analyzing the change in price rather than the price itself) to achieve stationarity. This paper demonstrates that this approach is often a trap. Differencing can induce spurious correlations between variables that were not present in the original system, leading to "ghost" edges in the resulting causal graph.
As shown in the figure above, differencing can create dependencies between error terms ($\Delta \epsilon$) that standard causal discovery tools interpret as direct causal links, even if the underlying relationship is purely lagged.
Methodology: Bridging VECM and SVAR-GFCI
To address these challenges, the authors combine an established causal discovery algorithm with a novel scoring metric designed for nonstationary data.
1. SVAR-GFCI Algorithm
The researchers build upon the "Structural Vector Autoregression - Greedy Fast Causal Inference" (SVAR-GFCI) algorithm. This hybrid approach utilizes both score-based search and independence tests to identify a Partial Ancestral Graph (PAG). PAGs are particularly useful because they can represent causal relationships even when some influential variables are unmeasured (latent confounders).
2. The VECM BIC Score
The core innovation is a modified Bayesian Information Criterion (BIC) score. Instead of a standard Vector Autoregression (VAR) model, which assumes stationarity, the authors employ a Vector Error-Correction Model (VECM). The VECM reparameterizes the system to account for both short-term changes (differences) and long-term equilibrium relationships (levels/cointegration). By calculating the likelihood based on the VECM, the BIC score remains a consistent criterion for model selection even when data is integrated of order 1 (I(1)).
3. Local Consistency
The authors prove that this new score is "locally consistent." This means that as the sample size increases, the score will correctly identify whether an edge should exist between two variables, effectively using the difference between BIC scores as a robust conditional independence test for nonstationary data.
Empirical Evaluation and Results
The method was evaluated using both synthetic simulations and a classic macroeconomic dataset.
Simulation Performance
The simulation involved a system with 5 observed variables and 2 latent variables designed as pure random walks. The results demonstrated that the VECM-based BIC score achieved significantly higher precision than standard methods. While a standard BIC score often became more "confident" in its spurious conclusions as sample size increased, the VECM-based approach correctly identified the underlying structure.
Insights from Danish Macroeconomic Data
When applied to quarterly data from Denmark (1973–2003), the algorithm discovered a feedback loop between inflation and money stock. It also suggested that two different interest rates were linked by a common latent mechanism. These findings align with established economic theories regarding market equilibrium but were derived purely from the data using the new algorithm.
Limitations and Future Directions
While this method represents a significant step forward, several areas for future research remain:
- I(1) Assumption: The current model specifically targets processes that become stationary after differencing once. It does not yet handle "explosive" processes or higher-order nonstationarity.
- Linearity: The approach assumes linear relationships, which may not capture the complexity of all economic systems.
- Explicit Cointegration Rank: The algorithm does not currently use the "cointegration rank" (the number of long-term equilibrium relationships) to guide its search. Incorporating this information could further refine the accuracy of the learned graphs.
Summary updated