Probabilistic Causal Impact takes aim at SHAP's blind spot

A causal explanation answers a narrow question: given what actually happened, which inputs made the difference? Principled answers to that question exist, and they have been computationally useless past the scale of a toy model. Producing one means walking through the counterfactual scenarios nobody observed.

A paper posted to arXiv on 3 September 2026 argues that the trade-off is a design choice, not a law. The framework, Probabilistic Causal Impact, or PCI, reframes explanation as a quantity to be estimated rather than a verdict computed exactly.

Two camps and the contradiction between them

As the abstract frames it, existing explainability tools split into two camps. Actual causality (AC) issues principled verdicts on which inputs deserve credit or blame, but only for toy-sized models, because computing a verdict means enumerating counterfactual scenarios. Scalable attribution methods such as SHAP, and even causal SHAP, cost far less, but they at least partially ignore the causal structure that generated the data. The consequence is what stings: those methods can return answers that conflict with a careful causal analysis. The same tension shows up wherever a model is asked to justify itself, as in explainable reinforcement learning for air traffic control.

PCI builds on actual causality and on Pearl's notions of probability of necessity and sufficiency, then reframes the problem as estimation on a probabilistic causal model that Monte Carlo sampling can approximate. Three pieces go in: a distribution over candidate explanations, a distribution over counterfactual values, and a scoring function. What comes out is a graded explanation rather than a yes-or-no verdict, with a probability attached to each candidate. Estimation as a way around an intractable exact computation has become the house style in machine learning, and the bar keeps dropping, as single-pass density and score estimation shows.

What the framework claims to add

The paper presents AC and Pearl's probability of causation as degenerate cases of PCI. That is the standard way a new framework argues it contains its predecessors instead of competing with them, and the argument only holds if the general case behaves. The general case is what the evaluation is meant to probe.

The most concrete item the abstract lists comes last: a real-world deployed causal machine learning model trained on millions of datapoints. Everything else on the list is a smaller exercise: synthetic and real-world examples, consistency checks against AC, scaling experiments, complex continuous-valued dynamical systems.

Where the evidence stops

No accuracy figure, runtime comparison, or baseline score appears in the abstract. For a paper whose entire pitch is computational feasibility, the missing number is the one reviewers and practitioners will open the full text to find. The listing carries no journal reference either, so as presented the framework's central claim is still an argument about design, not a demonstrated result. Benchmarks keep making the same point from the other direction: a model that recites an accounting formula is not the same as one that can apply it.

That is not a reason to dismiss it. It is a reason to read the claims as claims. The gap PCI targets is a real one, and the two camps it describes are not a strawman: the tension between principled causal verdicts on tiny models and fast attribution on large ones is a trade-off teams already live with in production.

Why the causal gap leaves the lab

Attribution scores travel further than their assumptions do. They end up in model cards, audit files, regulatory filings and postmortems, where the question being asked is usually causal even when the tool that produced the number is not: did this input cause the outcome, or did it merely move alongside it? An attribution method that ignores the structure of the data can produce a ranking that survives every internal review and still points a team at the wrong input to fix. The failure mode is not sloppiness but a correct calculation answering a question nobody asked, the same pattern behind automated statistical tests that land on the wrong conclusion.

What to watch for

Two things will decide whether PCI leaves the paper. The first is runtime on the continuous dynamical systems the abstract mentions, because Monte Carlo estimators trade accuracy for cost, and the paper has to show the trade favors it at the scale where AC falls over. The second is whether the deployed model in that evaluation is described in enough detail for anyone else to reproduce the result. The abstract does not name it, say what it predicts, or specify how PCI scored against it. Cost per unit of accuracy governs these decisions now, the way matching best-of-16 voting with 58.9% fewer tokens reframes the same question for routing.

A narrower takeaway holds either way for teams running attribution today. A SHAP-style ranking is a statement about the model's behavior under the data distribution. It is not a statement about cause, and the two can disagree even when both calculations are correct.