Abstract
Traditional causal discovery methods often assume that causal mechanisms are time-invariant. However, in dynamic systems like finance or neuroscience, these relationships frequently evolve. This paper introduces a framework that treats time as an explicit variable within functional causal models. By employing Gaussian Processes (GPs) to model time-varying coefficients, the authors enable the identification of non-stationary causal structures without the need for arbitrary windowing. Experiments on synthetic and real-world datasets, including stock indices and indoor temperatures, demonstrate that this approach accurately tracks evolving causal strengths and outperforms traditional sliding-window techniques.
Introduction
In causal discovery, the "rules of the game" are typically assumed to be fixed. If variable $A$ causes variable $B$ today, it is expected to do so identically tomorrow. Yet, in fields like neuroscience or economics, this assumption is often violated. Brain connectivity shifts depending on cognitive tasks, and economic relationships evolve with policy changes or market cycles. Applying static models to such non-stationary data often yields misleading results or "ghost" correlations.
Methodology
The authors extend the Functional Causal Model (FCM) by incorporating time $T$ as a root node that influences all observed variables. In this setup, an effect is a function of its causes, time, and independent noise: $$x_i(t) = f_i(t, \text{PA}_i) + \epsilon_i$$ where $\text{PA}_i$ represents the parents of $x_i$.
The Gaussian Process Approach
To avoid the pitfalls of "sliding window" methods—which are highly sensitive to the chosen window size—the authors use Gaussian Processes (GPs) as priors for the time-varying coefficients. GPs allow the model to automatically learn the smoothness of the relationship from the data. This approach also handles unobserved confounders by modeling them as time-dependent bias terms $g_i(t)$.
The estimation follows a two-step process:
- Lagged Effects: GP regression is used to estimate how past values influence current ones.
- Instantaneous Effects: Causal ordering at the same time step is determined by testing the independence of residuals using the Hilbert-Schmidt Independence Criterion (HSIC).
Experimental Results
The framework was validated using simulations and real-world datasets.
Simulations
In artificial time series with sinusoidal and step-function coefficients, the GP-based method significantly outperformed window-based methods. As shown in the figure below, the GP model tracks the ground truth coefficients much more closely than the discretized window approach.
Real-World Applications
- Stock Indices: Analyzing daily closing prices for the Nasdaq, FTSE, and N300 (2005–2014), the model correctly identified the sun-following causal flow (N300 → FTSE → Nasdaq). It also captured a dramatic spike in causal influence during the 2008 financial crisis, reflecting market contagion.
- House Temperature: In a dataset of room temperatures, the model detected reversals in causal direction when radiators were active and correctly identified the influence of unobserved confounders like human activity.
Conclusion
By treating time as a common cause and leveraging the flexibility of Gaussian Processes, this work provides a robust solution for identifying evolving causal structures. While computationally more intensive than static models, the ability to capture continuous dynamics without manual window tuning marks a significant advancement in causal discovery for non-stationary systems.
Summary updated