Causal Time Series
Summary
Causal structure appears as both a data-generation assumption and a reasoning task in the time-series cluster.
What The Wiki Currently Believes
- Causal-PIK is an active-intervention boundary case: a learned dynamics model predicts immediate action effects, an effect-space similarity becomes the covariance function of a GP reward surrogate, and GP-UCB plus simulation reranking chooses the next intervention. It is not causal discovery from observational time series, and the paper’s argument that entrywise-nonnegative similarity implies a PSD kernel is false in general.
- CauKer combines Gaussian-process kernel composition with structural causal models to generate synthetic, causally coherent time series.
- CaTSG explicitly defines observational, interventional, and counterfactual time-series generation and instantiates it with backdoor-adjusted diffusion guidance.
- TarDiff is not a causal model, but it is a useful warning source for healthcare generation: downstream clinical utility can diverge from average fidelity, while observational EHR data remain confounded logged decision data.
- TimeOmni-1 includes causality discovery as one of the perception capabilities in TSR-Suite.
- FactoryNet exposes setpoints/control inputs and measured feedback in industrial trajectories; its injected faults are events or experimental conditions, not automatically actions.
- FactoryBench turns the state/intervention/counterfactual/decision distinction into a public telemetry Q&A benchmark, but its L3 physical counterfactuals are approximate MMD-matched reruns rather than exact shared-state do-operator samples.
- Impact-Driven Event Embeddings is a terminology boundary case: its
impactis recurring exogenous-event response estimated from residuals, not an identified causal treatment effect or counterfactual intervention channel. - HUMI is a second boundary case: time-delayed mutual information can screen lagged event—observation dependence, but a peak does not identify causal direction under common causes, autocorrelation, calendar confounding, or selection effects.
- medDreamer is a real observational-EHR boundary case: it learns action-conditioned latent patient rollouts from logged treatments, but treatment assignment is confounded and its OPE/mortality proxies do not establish causal treatment-response counterfactuals.
- CFLight is a simulator-backed alternative-action boundary case: it makes exogenous-noise independence, monotonicity, and invertibility assumptions explicit, but fitting a BiCoGAN does not verify those assumptions or turn every generated branch into an identified physical counterfactual.
- CF-Reg is a terminology warning: its counterfactual is a prediction-flipping input perturbation around a classifier boundary, not an alternative action trajectory or causal intervention effect.
- Counterfactual PoNR Auditing is a simulator-identified boundary case: snapshot-matched branches identify intervention effects inside a fixed replayable simulator and audited capability class, but do not establish that the simulator captures real-world responses or that PoNR is intervention-class invariant.
Evidence
CauKer uses causality to create pretraining data; CaTSG uses causality to define generation targets; Causal-PIK uses dynamics-predicted immediate intervention effects to generalize reward observations across actions; TimeOmni-1 uses causality as a reasoning/evaluation target. Together they suggest causal structure is not optional if the goal is temporal understanding rather than curve fitting. TarDiff adds the utility caveat: even in healthcare, a generator can improve prediction without establishing causal intervention validity.
Relation To Foundation TSFM Agenda
This page maps directly to the causal/control slot in the Foundation Time-Series Model Research Agenda. Current local evidence is strongest for causal structure as synthetic-data prior or reasoning target; it is still weak for counterfactual action-conditioned rollout.
medDreamer sharpens that weakness rather than closing it. Explicit intervention channels and imagined futures are necessary interfaces for clinical world models, but observational action conditioning is not the same as causal identification; effective sample size, overlap, unmeasured confounding, reward-model error, and prospective transfer remain part of the evidence contract.
Tennessee Eastman Process Simulation Data adds an industrial boundary case: manipulated variables can be modeled as control-input channels, while fault injections are exogenous benchmark disturbances. This does not close counterfactual rollout without candidate interventions, rewards, or remediation actions.
Grid2Op adds a simulator-backed graph-control boundary case. In the early topology-controller challenge, injections are exogenous time series, topology is action-influenced, and overloads can feed back into topology through line disconnections and cooldown constraints. Later Grid2Op tracks add richer topology/control-input surfaces. This supports action-conditioned evaluation, but it is not evidence for causal discovery from observational time series or for a learned counterfactual rollout model.
Introducing machine learning for power system operation support adds a historical-labeling caveat: simulator replay of “what if this topology change had not occurred?” can extract plausible remedial-action labels from operator logs. This is counterfactual data construction, not proof that a learned model has identified causal structure.
LEAP nets for power grid perturbations sharpens the same split: synthetic topology variables can support controlled structural prediction, while the real RTE records use line-outage surrogates and should not be treated as causal identification from logged operator actions.
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Causal structure | partially closes | CauKer supplies causally structured synthetic generation; CaTSG partially closes counterfactual generation under a predefined SCM; TimeOmni-1 includes causal discovery as a reasoning task; HUMI adds a warning-quality lagged-dependence screen. | Needs transfer evidence on real temporal systems, conditional-dependence controls, interventions, and richer causal benchmarks. |
| Control and counterfactuals | partially closes | CaTSG derives interventional and counterfactual diffusion objectives; Causal-PIK uses short action-conditioned effect prediction for sequential intervention selection; Tennessee Eastman and FactoryNet expose control-input channels; L2RPN/Grid2Op exposes simulator-backed topology control; FactoryBench tests intervention and approximate physical-counterfactual questions. | Needs long-horizon learned candidate-action rollout, exact intervention semantics, calibrated alternative outcomes, and causal identification beyond simulator-defined, matched-rerun, or synthetic transitions. |
| Recoverability auditing | partially closes | CPA supplies snapshot-matched finite-horizon recoverability labels and a deployed-versus-upper-bound policy-gap decomposition inside a replayable simulator. | Needs real-system validation, wider intervention families, model-error-aware bounds, hidden-state treatment, and comparison with viability/reachability methods. |
Open Questions
- How much causal correctness is needed for synthetic pretraining to transfer?
- Can models learn causality from synthetic templates without overfitting to template artifacts?
- How should real-world counterfactual generation be evaluated when ground-truth counterfactuals are unavailable?