DCD: Decomposition-based Causal Discovery from Autocorrelated and Non-Stationary Temporal Data
Pith reviewed 2026-05-16 08:29 UTC · model grok-4.3
The pith
Decomposing each time series into trend, seasonal, and residual parts recovers causal structure more accurately than raw-data methods
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The DCD framework decomposes each time series into trend, seasonal, and residual components, applies stationarity tests to the trend, kernel-based dependence measures to the seasonal part, and constraint-based causal discovery to the residual, then integrates the resulting component-level graphs into a single multi-scale causal structure that isolates long- and short-range effects and reduces spurious associations.
What carries the argument
Decomposition of each time series into trend, seasonal, and residual components, with component-specific causal inference followed by integration into a unified multi-scale graph.
If this is right
- More accurate recovery of ground-truth causal structure than state-of-the-art baselines on synthetic benchmarks
- Particularly strong performance under strong non-stationarity and temporal autocorrelation
- Reduced spurious edges and improved interpretability of the recovered graph
- Demonstrated gains on real-world climate time series
Where Pith is reading between the lines
- The same decomposition step could be tested on economic or physiological series that also contain strong cycles
- The multi-scale output graph might support separate modeling of slow versus fast causal influences
- Alternative decomposition families, such as wavelets, could be swapped in to check robustness
Load-bearing premise
The chosen decomposition isolates the true causal effects without introducing artifacts or discarding critical dependency information.
What would settle it
On synthetic series where the decomposition is known to mix causal signals, the method would need to recover ground-truth structure worse than or equal to raw-data baselines.
read the original abstract
Multivariate time series in domains such as finance, climate science, and healthcare often exhibit long-term trends, seasonal patterns, and short-term fluctuations, complicating causal inference under non-stationarity and autocorrelation. Existing causal discovery methods typically operate on raw observations, making them vulnerable to spurious edges and misattributed temporal dependencies. We introduce a decomposition-based causal discovery framework that separates each time series into trend, seasonal, and residual components and performs component-specific causal analysis. Trend components are assessed using stationarity tests, seasonal components using kernel-based dependence measures, and residual components using constraint-based causal discovery. The resulting component-level graphs are integrated into a unified multi-scale causal structure. This approach isolates long- and short-range causal effects, reduces spurious associations, and improves interpretability. Across extensive synthetic benchmarks and real-world climate data, our framework more accurately recovers ground-truth causal structure than state-of-the-art baselines, particularly under strong non-stationarity and temporal autocorrelation.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript introduces DCD, a decomposition-based causal discovery framework for multivariate time series with long-term trends, seasonal patterns, and short-term fluctuations under non-stationarity and autocorrelation. Each series is separated into trend, seasonal, and residual components; trends are analyzed via stationarity tests, seasonal components via kernel-based dependence measures, and residuals via constraint-based causal discovery. Component-level graphs are then integrated into a unified multi-scale causal structure. The authors claim this isolates long- and short-range effects, reduces spurious associations, and outperforms state-of-the-art baselines on extensive synthetic benchmarks and real-world climate data.
Significance. If the central claim holds, the framework could meaningfully advance causal discovery for non-stationary temporal data by explicitly handling multi-scale structure, with potential impact in climate science and related domains. The reported empirical gains on benchmarks under strong autocorrelation and non-stationarity would constitute a useful practical contribution, especially if the method avoids the artifacts that plague raw-data approaches.
major comments (3)
- [Abstract] Abstract: The decomposition method itself is unspecified (e.g., STL, X-11, or another procedure), yet the central claim that it 'isolates causal effects without introducing artifacts' is load-bearing. Standard decompositions can split a single causal link across components or induce spurious residual correlations in autocorrelated non-stationary series; this risk must be analyzed or empirically bounded in the methods section.
- [Methods] Methods (integration step): No description is given of how component-level graphs are merged when cross-component dependencies or conflicts arise. This step is required to produce the claimed 'unified multi-scale causal structure' and must be formalized, including any assumptions about independence across scales.
- [Experiments] Experiments (synthetic benchmarks): The data-generation process must be shown not to favor the proposed decomposition (e.g., by defining ground-truth edges on the raw series rather than post-decomposition). If the benchmarks are consistent with data generated to match the decomposition, the reported superiority does not yet demonstrate general robustness under the skeptic's concern.
minor comments (1)
- [Abstract] Abstract: Quantitative metrics (e.g., average SHD or F1-score deltas versus baselines) should be stated to support the performance claim rather than the qualitative statement 'more accurately recovers'.
Simulated Author's Rebuttal
We thank the referee for their constructive and detailed comments, which have helped us identify areas where the manuscript can be strengthened for clarity and rigor. We address each major comment below and outline the corresponding revisions.
read point-by-point responses
-
Referee: [Abstract] Abstract: The decomposition method itself is unspecified (e.g., STL, X-11, or another procedure), yet the central claim that it 'isolates causal effects without introducing artifacts' is load-bearing. Standard decompositions can split a single causal link across components or induce spurious residual correlations in autocorrelated non-stationary series; this risk must be analyzed or empirically bounded in the methods section.
Authors: We agree that the specific decomposition procedure and its potential artifacts require explicit treatment. The current manuscript describes the overall framework but does not name the decomposition algorithm or provide bounds on induced correlations. In the revised version we will (i) state in the abstract and Section 3 that STL decomposition is used with fixed parameters, (ii) add a dedicated subsection analyzing how decomposition can split or create spurious links, and (iii) include both theoretical discussion and new empirical bounds on residual correlations under controlled autocorrelation levels. revision: yes
-
Referee: [Methods] Methods (integration step): No description is given of how component-level graphs are merged when cross-component dependencies or conflicts arise. This step is required to produce the claimed 'unified multi-scale causal structure' and must be formalized, including any assumptions about independence across scales.
Authors: We acknowledge that the integration step is currently described only at a high level. We will add a formal subsection (new Section 3.4) that defines the merging operator, specifies the conflict-resolution rule (union of edges with tie-breaking by component-wise p-value), provides pseudocode, and states the modeling assumption that causal influences at different scales are conditionally independent given the observed series. This will make the construction of the unified multi-scale graph fully reproducible. revision: yes
-
Referee: [Experiments] Experiments (synthetic benchmarks): The data-generation process must be shown not to favor the proposed decomposition (e.g., by defining ground-truth edges on the raw series rather than post-decomposition). If the benchmarks are consistent with data generated to match the decomposition, the reported superiority does not yet demonstrate general robustness under the skeptic's concern.
Authors: The ground-truth edges in our synthetic benchmarks are defined on the raw series via the structural equations before any decomposition is applied (Section 4.1). Non-stationarity and autocorrelation are generated independently of the decomposition routine. Nevertheless, to directly address the referee’s concern we will add (i) an explicit statement confirming the pre-decomposition definition of ground truth and (ii) a new set of experiments that re-verify recovered edges against the raw-series ground truth after decomposition. These additions will be presented as a robustness check. revision: partial
Circularity Check
No circularity: method applies established decomposition and discovery tools to components
full rationale
The derivation chain decomposes series via standard methods, applies stationarity tests to trends, kernel dependence to seasonals, and constraint-based discovery to residuals, then integrates graphs. No step reduces by construction to its own inputs, no fitted parameter is relabeled as prediction, and no load-bearing claim rests on self-citation chains or imported uniqueness theorems. Benchmarks compare against baselines on synthetic and climate data; the result is not equivalent to the decomposition inputs. This is the common honest case of an applied framework whose central claim remains independently testable.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Multivariate time series can be decomposed into trend, seasonal, and residual components that preserve causal relationships.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.