Prediction Bottlenecks Don't Discover Causal Structure (But Here's What They Actually Do)

Technical Summary: Prediction Bottlenecks Don't Discover Causal Structure (But Here's What They Actually Do)

1. Problem Statement

The paper investigates a specific hypothesis in time-series modeling: whether a forecaster trained solely on next-step prediction (minimizing Mean Squared Error) implicitly learns causal or dependency structures that can be extracted from its weights. This hypothesis suggests that "causal discovery" could emerge as a byproduct of strong forecasting, potentially turning every pretrained forecaster into a causal-discovery engine.

The authors focus on a specific claim involving Mamba state-space models (SSMs). Previous observations suggested that for a Mamba model trained on next-step MSE, the product of its output and input projection matrices, $S = |W_{out}W_{in}|$, reliably recovers Granger-causal graphs on synthetic benchmarks. Furthermore, initial experiments indicated this effect generalized across architectures and was significantly amplified by interventional data ($p < 10^{-5}$).

The paper's objective is to rigorously test these claims using a standardized, reusable falsification benchmark to determine if the phenomenon is a genuine causal discovery mechanism or an artifact of experimental design.

2. Methodology: The Falsification Benchmark

The authors constructed a five-stage protocol designed to stress-test the extraction method along axes the original claim had ignored. This protocol serves as a reusable benchmark for future "method-discovers-causality" papers.

The Extraction Protocol:
For a $K$-variable time series $X \in \mathbb{R}^{T \times K}$, a model is trained with input projection $W_{in}$ and output projection $W_{out}$. After training, the causal graph is extracted via $S = |W_{out}W_{in}|$, where $S_{i,j}$ represents the strength of $j \to i$.

The Five Stress Regimes (Falsifications):

1. Architecture/Capacity (F1): Comparing the SSM against five other architectures (Linear, Transformer, LSTM, MLP) with matched capacity across six synthetic generators.
2. Sparsity and Dimensionality (F2): A focused stress grid over $K \in \{10, 20\}$, $T \in \{150, 300\}$, and max-lag 8, including nonlinear sweeps.
3. Real-World Ground Truth (F3): Evaluation on three datasets with varying ground-truth quality:
   • Climate (NOAA): Soft labels (3 edges, $K=6$).
   • Finance (ETFs): Soft labels (6 edges, $K=10$).
   • Lorenz-96: Clean, unambiguous ground truth (90 edges, $K=10$).
   • Note: Each dataset includes an "edge-provenance card" to audit the validity of ground-truth labels.
4. Intervention Semantics (F4): Testing three intervention types: standard $do(X=c)$ clamps, soft additive noise, and non-standard random forcing. Crucially, this includes size-matched control arms to prevent conflating intervention content with increased sample size.
5. Method-Agnosticity (F5): Applying the same intervention protocol to classical baselines (Lasso, PCMCI, Bivariate Granger) to see if the observed effects are unique to the bottleneck or method-agnostic.

3. Key Results and Falsifications

The study systematically dismantles the original claims through five specific falsifications:

• F1: Architecture Specificity Fails. A plain linear bottleneck matches or outperforms the Mamba SSM on every synthetic dataset when capacity and seeds are matched. The "emergence" of causal structure is attributed to the linear bottleneck performing low-rank lag-1 regression, which the SSM replicates at a higher parameter cost.
• F2: Competitiveness with Sparse Regression Fails. On synthetic CauseMe-style benchmarks and the Lorenz-96 dataset, tuned Lasso and classical methods (PCMCI, Granger) significantly outperform the bottleneck. On the Lorenz-96 benchmark (the only one with unambiguous ground truth), the bottleneck trails the tight cluster of classical and modern causal baselines.
• F3: Real-World Performance is Unstable. On the Lorenz-96 benchmark, the bottleneck performs poorly compared to classical methods. On the climate dataset, rankings are highly sensitive to ground-truth definitions. When "definitional" edges (e.g., ENSO and SOI, which are physically the same signal with opposite signs) are removed from the ground truth, the ranking of methods flips entirely, demonstrating that previous "wins" were often artifacts of capturing definitional correlations rather than causal discovery.
• F4: The Intervention Advantage is a Confound. The headline claim that interventional data boosts performance ($p < 10^{-5}$) collapses under size-matched controls. The original effect was largely driven by the intervention arm having more total samples ($T + K \cdot T_{int}$) than the observational arm. When sample sizes are matched, the gain shrinks to a negligible amount ($+0.03$ to $+0.05$). Furthermore, the remaining gain only survives under a non-standard "random-forcing" scheme, not standard $do(X=c)$ interventions.
• F5: The Effect is Method-Agnostic. The residual gain observed under random forcing is not unique to the bottleneck. Classical Bivariate Granger and PCMCI show the same or larger effects under the same size-matched controls. The effect is best explained as robustness to target corruption (where shared weights absorb noise better than per-equation regression like Lasso) rather than a specific ability to extract causal structure.

4. What Survives (Characterization Claims)

After falsifying the causal discovery claims, the authors identify three narrow, valid characterization findings:

1. Mild-Nonlinearity Configuration: In a specific configuration ($K=20, T=300$) with mild nonlinearity, the bottleneck beats tuned baselines on 87% of trials. This is a specific configuration, not a general regime.
2. Sample Efficiency: The bottleneck gains slightly more from additional observational data than Lasso does, which may be useful where collecting data is cheap.
3. Target-Corruption Robustness: The bottleneck's shared output weights ($W_{out}$) are more robust to per-step random forcing (corrupted target rows) than Lasso's per-equation fits. This explains the residual intervention effect but is a reliability result, not a causal discovery result.

5. Significance and Contributions

The paper's primary contribution is negative: it demonstrates that the proposed method does not discover causal structure as claimed. The significance lies in:

• A Reusable Falsification Benchmark: The authors release a standardized suite (synthetic generators, intervention semantics, edge-provenance cards, and control arms) to prevent future papers from making similar unsubstantiated claims.
• Methodological Lessons:
  • Size-matched controls are critical: Increasing seed counts cannot diagnose sample-size confounds; structural controls are required.
  • Intervention semantics matter: The difference between "random forcing" and proper $do(X=c)$ interventions is the difference between a significant effect and a null result.
  • Prediction baselines $\neq$ Causal baselines: Future claims must include classical causal methods (e.g., Granger, PCMCI) in intervention control arms, as they often benefit from the same mechanisms.
  • Ground-truth auditing: Rankings on soft benchmarks are highly sensitive to definitional couplings in the ground truth (e.g., ENSO/SOI).

Conclusion: The paper concludes that the answer to whether a free causal-discovery method is hidden inside prediction bottlenecks is no. The headline effects fail under proper controls or reproduce as larger effects in classical methods. The lasting artifact of this work is the benchmark itself and the narrow characterization of what the bottleneck actually achieves (robustness and mild nonlinearity handling) rather than causal discovery.