Granger Causality Networks

• Granger causality networks are directed graphs where an edge from series i to j indicates that i's past improves the forecast of j's future.
• They are estimated using methods like linear VAR with sparsity-inducing penalties or neural models that capture nonlinear dependencies.
• These networks address challenges such as confounding, nonstationarity, and model misspecification for accurate prediction and inference.

A Granger causality network is a directed graph over a multivariate time series in which an edge $i \to j$ indicates that the past of series $i$ improves prediction of the future of series $j$ beyond what can be achieved from the past of the remaining variables. In the standard vector autoregressive setting, this network is encoded by the nonzero pattern of lag matrices; in nonlinear settings it is encoded by the dependence structure of component-wise predictive functions. The modern literature treats the Granger causality network as a time-series graphical object that can be estimated in linear, nonlinear, high-dimensional, nonstationary, and partially confounded regimes, while retaining the predictive definition of causality (Shojaie et al., 2021, Liu et al., 15 Jan 2025).

1. Formal definition and graph-theoretic representation

For a multivariate process $x_t = (x_{1t},\ldots,x_{pt})^\top$, Granger’s predictive definition can be stated as a comparison of conditional prediction errors: series $y$ is Granger-causal for series $x$ if including the past of $y$ reduces the prediction error variance relative to prediction from the remaining history alone (Shojaie et al., 2021). In the linear VAR($d$) model,

$x_t = \sum_{k=1}^d A^k x_{t-k} + e_t,$

series $x_i$ is Granger-causal for series $i$0 if and only if $i$1 for some lag $i$2 (Shojaie et al., 2021). This gives the classical network rule: nodes are time series, and a directed edge $i$3 exists when at least one lag coefficient from $i$4 to $i$5 is nonzero.

The literature distinguishes pairwise and conditional Granger causality. Pairwise or bivariate Granger causality tests whether one series helps predict another using only the past of those two series; conditional or multivariate Granger causality tests whether a series helps predict a target conditional on the past of all other relevant series (Shojaie et al., 2021). For network construction, the conditional notion is the more faithful one, because omitted variables can create spurious links or reverse apparent directions.

Graphically, two related representations appear in the VAR literature. The expanded graph replicates nodes across time slices and places an edge from $i$6 to $i$7 when $i$8; the compact Granger causal graph collapses across lags and places a single directed edge $i$9 whenever some lag coefficient is nonzero (Shojaie et al., 2021). In the compact graph, undirected edges can also encode instantaneous dependence through nonzero entries of $j$0, where $j$1 is the innovation covariance (Shojaie et al., 2021). This places Granger causality networks in close relation to dynamic graphical models.

The same graph concept extends beyond linearity. In a nonlinear autoregressive formulation,

$j$2

series $j$3 is Granger-noncausal for $j$4 if $j$5 is invariant to the history of $j$6 (Shojaie et al., 2021, Yin et al., 2022). The network then records which component-wise predictive functions depend on which histories.

2. Linear estimation, sparsity, and high-dimensional network recovery

In high-dimensional settings, naive VAR estimation becomes unstable because the number of parameters grows as $j$7 (Shojaie et al., 2021, Aslan et al., 2024). A standard remedy is penalized least squares,

$j$8

with $j$9 chosen to induce sparsity in the network (Shojaie et al., 2021). Entrywise lasso encourages individual coefficients to vanish, whereas group lasso over the lag vector $x_t = (x_{1t},\ldots,x_{pt})^\top$0 aligns sparsity with the existence of a directed edge $x_t = (x_{1t},\ldots,x_{pt})^\top$1 (Shojaie et al., 2021).

A particularly explicit formulation appears in thresholded group-lasso network Granger causality. There, variables are partitioned into non-overlapping groups, and estimation uses a group-lasso penalty on lagged blocks $x_t = (x_{1t},\ldots,x_{pt})^\top$2 so that entire groups of candidate parents can be selected or discarded together (Basu et al., 2012). The thresholded variant applies group-level thresholding and then within-group thresholding, with the stated goal of handling group misspecification and improving support recovery (Basu et al., 2012). This yields a network whose directed edges are read off from nonzero lagged coefficients after thresholding.

High-dimensional linear work also includes lag-selection penalties such as truncating lasso and hierarchical group lasso, joint estimation of directed lagged edges and instantaneous undirected edges through sparse $x_t = (x_{1t},\ldots,x_{pt})^\top$3, and sparse-plus-low-rank decompositions $x_t = (x_{1t},\ldots,x_{pt})^\top$4 to separate direct interactions from latent common drivers (Shojaie et al., 2021). Factor-augmented formulations serve the same purpose when interest centers on a subset of series but a large set of additional variables must be controlled (Shojaie et al., 2021).

These methods establish the classical interpretation of a Granger causality network as a sparse directed graph estimated from zero and nonzero coefficient structure. Their main limitation is model misspecification under nonlinear dynamics.

3. Nonlinear and neural Granger causality networks

Modern nonlinear approaches replace the linear predictor by a flexible function approximator but preserve the same network logic: a directed edge exists when the predictor for one component depends on the lagged history of another component. A widely used template is the component-wise neural autoregression, where one network is trained per target variable and sparsity is imposed on parameters that connect source histories to the target predictor (Shojaie et al., 2021).

One influential extension is the self-explaining neural network formulation of generalized vector autoregression. In GVAR,

$x_t = (x_{1t},\ldots,x_{pt})^\top$5

the matrices $x_t = (x_{1t},\ldots,x_{pt})^\top$6 are state-dependent generalized coefficients rather than fixed VAR coefficients (Marcinkevičs et al., 2021). This yields a Granger causality network whose edge strengths and signs can vary over time. GVAR summarizes an edge $x_t = (x_{1t},\ldots,x_{pt})^\top$7 by aggregating the absolute generalized coefficients across time and lags, and then uses a stability-based thresholding procedure with time-reversed data to select a binary adjacency matrix (Marcinkevičs et al., 2021).

Another line of work decouples series selection from lag selection. In the decoupled-filter framework, a series-level filter $x_t = (x_{1t},\ldots,x_{pt})^\top$8 gates whether source series $x_t = (x_{1t},\ldots,x_{pt})^\top$9 influences target $y$0, while a lag-level filter $y$1 selects which lags of that source remain active (Sultan et al., 2022). This construction is architecture-agnostic and can sit in front of MLPs, RNNs, LSTMs, or Transformers, making the Granger network readable directly from explicit filter variables rather than internal weights (Sultan et al., 2022). The same work introduces LeKVAR, which preserves VAR-style edge interpretation through linear matrices $y$2 after a learned element-wise nonlinear kernel transformation (Sultan et al., 2022).

Jacobian-based approaches shift the focus from sparse weights to input-output sensitivity. Jacobian Granger Causality defines the importance of lagged input $y$3 for predicting $y$4 through the derivative $y$5, aggregates these Jacobians over time, and thresholds the resulting variable-lag scores to infer a directed graph with lags and signs (Suryadi et al., 2022). JRNGC generalizes this to a single multivariate predictor for all targets and regularizes the input-output Jacobian directly, using either an $y$6 or Frobenius-norm penalty to obtain both summary and full-time Granger causality networks (Zhou et al., 2024). In these models, the network is a post-hoc object derived from sensitivity structure rather than first-layer sparsity.

Kolmogorov–Arnold network approaches use learned univariate functions on edges. GC-KAN learns one KAN per target series, applies sparsity-promoting regularization to first-layer edge functions, and reads lag-specific influence from the magnitudes of those edge functions (Lin et al., 2024). KANGCI, also called Granger Causal KAN, uses component-wise nonlinear autoregressions, group lasso on the columns of the first-layer base-weight matrix $y$7, ridge regularization on deeper base weights, and a time-reversed Granger causality algorithm that either selects the original or reversed adjacency matrix or fuses them elementwise (Liu et al., 15 Jan 2025). In both formulations, the Granger causality network is extracted from first-layer KAN parameters, yielding a directed weighted graph over the observed variables.

A different neural paradigm treats Granger causality not as sparse regression but as predictive degradation under variable removal. A joint deep model is trained to predict the entire multivariate future, and Granger structure is then inferred by comparing residual distributions or model uncertainty between a full input and a reduced input in which one source series is masked (Sultan et al., 25 Jun 2025). This preserves the classical restricted-versus-full model logic without explicit sparsity penalties.

4. Confounding, nonstationarity, and time-resolved networks

A central difficulty in Granger network estimation is confounding. Classical Granger causality assumes causal sufficiency; unobserved common drivers can create spurious predictive links (Shojaie et al., 2021). One explicit neural response is deep recurrent modelling with latent confounding, which introduces a latent confounder $y$8, noisy proxies $y$9, a proxy encoder GRU, and a dual-decoder architecture that compares restricted and full predictive distributions for the target while conditioning on an inferred substitute confounder $x$0 (Yin et al., 2022). In this framework, a directed edge is not read from sparse weights but from statistically significant reductions in prediction error when the candidate cause is added, conditional on the inferred confounder representation (Yin et al., 2022).

High-dimensional background adjustment can also be performed without fitting a full multivariate VAR. In networks with a small number of nodes of interest and a large number of additional nodes, spectral dynamic principal component analysis summarizes the background network through dynamic principal component scores, regresses these scores out of the nodes of interest, and then applies standard bivariate Granger causality to the residual series (Aslan et al., 2024). The resulting network is a background-adjusted Granger causality network rather than a naive pairwise graph.

Nonstationarity introduces a different challenge. Sliding-window multivariate Granger causality has been used to study dynamic effective connectivity in a multistable biological decision network model, where pairwise conditional GC is computed in 75–400 ms windows and the inferred directed connectivity varies across stimulus-locked versus decision-locked analyses and across correct versus error trials (Asadpour et al., 2024). Jacobian-based neural methods provide another route: by augmenting the predictor with an explicit time variable, JGC produces Jacobian scores $x$1 that become functions of time and can reveal regime changes or structural breaks in a time-varying Granger network (Suryadi et al., 2022).

These developments make clear that a Granger causality network need not be static. Depending on the model class, it may be a summary graph, a lag-resolved graph, or a full-time sequence of graphs indexed by time windows or by explicit time-varying parameters.

5. Empirical domains and characteristic network structures

Granger causality networks are used across neuroscience, genomics, finance, climate-related dynamics, and symbolic event streams. In the KANGCI evaluation, the network is recovered from Lorenz-96, DREAM-3 and DREAM-4 gene regulatory networks, simulated fMRI BOLD signals, synthetic VAR systems, and real-world EEG datasets, with the final object of interest being the adjacency matrix $x$2 whose entries quantify nonlinear Granger influence $x$3 (Liu et al., 15 Jan 2025). The same paper interprets Lorenz-96 as near-nearest-neighbor ring-like connectivity, gene networks as sparse directed interactions, and fMRI networks as connectivity patterns tied to the underlying DCM and balloon-model structure (Liu et al., 15 Jan 2025).

In neuroscience, high-dimensional EEG networks motivate both background-adjusted and fully multivariate approaches. The sDPCA-based method has been applied to multichannel EEG with nodes of interest such as AF3, AF4, C3, C4, O1, and O2, producing directed connectivity patterns that differ between motor imagery and motor execution (Aslan et al., 2024). Pairwise conditional GC in a multistable decision network reveals directed interactions among sensorimotor, inhibitory, uncertainty, and motor units that vary with task difficulty and with correct versus error decisions (Asadpour et al., 2024). In latent-confounder modelling, the Danube river discharge example shows how the same predictive logic can also be used in small systems with proxies for common causes, where the model correctly infers no lagged Granger causality from Iller at Kempten to Danube at Dillingen (Yin et al., 2022).

Financial applications often emphasize the network as a dynamic system of return spillovers. In a study of 20 developed stock markets from 2 January 2006 to 31 December 2013, daily closing prices were aligned to account for non-synchronous trading, 94 rolling sub-sample networks were constructed, and spatial probit models were used to study determinants of directed edges (Výrost et al., 2014). That analysis reports three substantive findings: after the recent financial crisis the impact of the US stock market declined; temporal proximity between market closing times matters for return spillovers; and preferential attachment exists, meaning that spillover from market $x$4 to market $x$5 is more likely when $x$6 influences other markets or $x$7 is influenced by other markets (Výrost et al., 2014).

Symbolic and quantized data admit a different but related network formalism. For symbolic streams generated by ergodic stationary sources, causal dependence can be represented by directed edges weighted by a coefficient $x$8 and explicit crossed probabilistic automata, yielding a causal network over symbolic processes rather than continuous-valued variables (Chattopadhyay, 2014). This extends the idea of a Granger causality network beyond VARs and neural autoregressions to automata-theoretic predictive models.

6. Interpretation, common misconceptions, and limitations

A recurrent misconception is that a Granger causality network is automatically a structural causal graph in the interventionist sense. The review literature is explicit that Granger causality is not causal effect in the Pearl or Holland sense; it is a model-based notion of directed predictive dependence under temporal ordering and conditioning assumptions (Shojaie et al., 2021). Accordingly, an edge $x$9 means that the past of $y$0 improves prediction of $y$1 in the chosen model class, not that an intervention on $y$2 has been identified.

Another common misconception is that pairwise networks are sufficient. In fact, omitted variables, latent confounders, and indirect paths can induce spurious edges or even direction reversals (Shojaie et al., 2021, Yin et al., 2022). This is why conditional GC, proxy-based confounder adjustment, sparse-plus-low-rank decompositions, factor adjustment, or network-adjusted residualization appear repeatedly in the modern literature.

Model dependence is equally central. Linear VAR networks can be badly misspecified under nonlinear dynamics; nonlinear networks can overfit in small samples; time-reversal heuristics have strong theoretical support mainly for linear finite-order autoregressive systems and are more nuanced in nonlinear chaotic settings (Liu et al., 15 Jan 2025). Hyperparameters such as lag order, sparsity strength, network width, spline granularity, or Jacobian regularization weight materially affect the inferred graph (Zhou et al., 2024, Liu et al., 15 Jan 2025). In ultra-dense or very strongly connected systems, low-dimensional adjustments may fail to remove enough confounding variation, and accuracy can stabilize at substantially lower levels than in sparse settings (Aslan et al., 2024).

Sampling and stationarity assumptions remain fundamental. The review literature emphasizes that subsampling and mixed-frequency observations can generate missing or spurious Granger edges if the sampling structure is ignored (Shojaie et al., 2021). Nonstationarity can make a single static graph misleading; in such settings, sliding-window GC, regime-specific models, or time-augmented predictors are more appropriate (Asadpour et al., 2024, Suryadi et al., 2022).

Taken together, these caveats do not diminish the utility of Granger causality networks. They specify the object more precisely: a directed graph of predictive influence whose interpretation depends on the conditioning set, sampling regime, model class, and regularization scheme. Within those constraints, the Granger causality network remains one of the most technically developed representations of directed dependence in multivariate time series.