Instantaneous Causal Discovery for High Frequency Time Series

This direction is merged into another direction: temporal resolution mismatch. Its last version stays readable and citable here.

Version 1 of 163 papers78 conceptsfrozen 2026-10-04

In the study of time series, the gold standard for finding "who causes what" has long been temporal precedence: if event $A$ consistently happens before event $B$, $A$ might be the cause. This is the intuition behind Granger causality and standard Dynamic Bayesian Networks. However, this logic breaks down when the underlying causal mechanism moves faster than our sensors can record.

When a biological signal, a stock trade, or a physical interaction occurs at a higher frequency than the data sampling rate, the cause and its effect appear in the same discrete time step. To a standard model, these look like simultaneous correlations, leaving the direction of influence ambiguous. Researchers are now developing frameworks to "see between the frames," identifying instantaneous causal effects that occur within a single observation window.

The Sampling Rate Gap

The fundamental problem is a mismatch between physical reality and data discretization. As Dirk Husmeier, Frank Dondelinger, and Sophie Lèbre point out in Inter-time segment information sharing for non-homogeneous dynamic Bayesian networks, standard models assume a fixed time lag. If a feedback loop is instantaneous or very fast, these models fail to capture the transition.

Imagine a video of a billiard ball hitting another. If the camera captures 60 frames per second, we see the collision and the subsequent movement. But if the camera only captures one frame every ten seconds, we might just see two balls in new positions simultaneously. Yuxiao Cheng and colleagues argue in CUTS+: High-Dimensional Causal Discovery from Irregular Time-Series that when mechanisms operate faster than the collection frequency, causal relationships become indistinguishable from non-causal correlations unless we explicitly model contemporaneous effects.

Breaking Symmetry with Invariance and Nonstationarity

If $A$ and $B$ happen at the same time, how do we know $A \to B$ instead of $B \to A$? The field has moved toward using "distribution shifts" as a diagnostic tool.

Niklas Pfister, Peter Bühlmann, and Jonas Peters pioneered this in Invariant Causal Prediction for Sequential Data. Their intuition is that a true causal relationship $P(Effect | Cause)$ should remain invariant even when the environment changes. If the relationship between $A$ and $B$ stays the same across different time segments or "interventions," but the marginal distribution of $A$ changes, then $A$ is likely the cause. This allows for the detection of instantaneous links that Granger causality—which relies strictly on lags—would miss.

Similarly, Weiran Yao and Kun Zhang have shown in Temporally Disentangled Representation Learning that nonstationarity (the way data distributions shift over time) is actually a source of information. By observing how variables change together across different "contexts," we can identify latent causal variables and their intra-step interactions without needing a time delay.

Causal Representation Learning

A significant hurdle in this research is that we often don't observe the causal variables directly; we observe high-dimensional proxies like video pixels or sensor arrays.

The iCITRIS framework, developed by Phillip Lippe and Sara Magliacane in iCITRIS: Causal Representation Learning for Instantaneous Temporal Effects, tackles this by using known interventions (like a robot's actions) to resolve the "chicken-and-egg" problem. If we know we moved a specific object, we can see which other variables change in the same frame, allowing the model to distinguish between the causal variable and its instantaneous effects.

For systems where interventions aren't available, Yifan Shen and Ruichu Cai introduced IDOL in On the Identification of Temporally Causal Representation with Instantaneous Dependence. IDOL works by assuming that causal relationships are sparse and leveraging the temporal context of surrounding points to "triangulate" the direction of instantaneous influence.

Continuous Time as a Solution

Rather than forcing data into discrete "buckets" of time, some researchers suggest modeling the system as a continuous flow. Benjie Wang and Wenbo Gong introduced SCOTCH in Neural Structure Learning with Stochastic Differential Equations. By using Neural SDEs, they treat the causal graph as a set of continuous interactions. This naturally handles irregularly sampled data and avoids the arbitrary "instantaneous vs. lagged" distinction by allowing influences to propagate at any speed, even those much faster than the observation interval.

Why It Matters

The shift toward modeling instantaneous effects is not just a mathematical curiosity; it is essential for robustness. As demonstrated by Sumanth Varambally and Rose Yu in Discovering Mixtures of Structural Causal Models from Time Series Data, accounting for both lagged and simultaneous links provides a far more accurate structural representation of complex systems like financial markets or gene networks. Without these contemporaneous links, models are forced to treat real causal influences as "noise" or "unobserved confounding," leading to poor generalizations and failed interventions.

Papers in this version

Showing 60 of 63 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_instantaneous_causal_discovery_for_high__3e4ee035,
  title = {Instantaneous Causal Discovery for High Frequency Time Series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/instantaneous-causal-discovery-for-high-frequency-time-series/rdr_3e4ee03546d74d25b9aafe18856cfae8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_instantaneous_causal_discovery_for_high__3e4ee035_v1,
  title = {Instantaneous Causal Discovery for High Frequency Time Series},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/instantaneous-causal-discovery-for-high-frequency-time-series/rdr_3e4ee03546d74d25b9aafe18856cfae8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}