Financial Causal Attribution Score (FCAS)

• FCAS is a metric that quantifies alignment between a model’s predicted stock return direction and the causal polarity extracted from financial news texts.
• It is computed by comparing the sign of a model’s prediction with the extracted causal cue to diagnose narrative consistency across different macroeconomic regimes.
• Its application highlights model robustness under semantic shifts, though scaling inconsistencies and extraction ambiguities pose reproducibility challenges.

Searching arXiv for the primary paper and closely related references to ground the article in current arXiv records. Search 1: primary paper on FCAS. Search 2: causal stress testing and causal-effect score papers for related context. Financial Causal Attribution Score (FCAS) is a metric introduced in the regime-aware evaluation framework of “Quantifying Semantic Shift in Financial NLP: Robust Metrics for Market Prediction Stability” to measure whether a model’s predicted stock return direction agrees with causal claims extracted from financial news text. Within that framework, FCAS is one of four diagnostics—alongside Patent Cliff Sensitivity (PCS), Temporal Semantic Volatility (TSV), and NLI-based Logical Consistency Score (NLICS)—used to assess robustness of financial NLP models under macroeconomic regime shift, with the stated aims of supporting auditability, stress testing, and adaptive retraining in financial AI systems (Sun et al., 30 Sep 2025).

1. Definition and intended scope

FCAS is defined as a measure of alignment between a model’s predicted return direction and the causal signal implied by a news article. In the paper’s prose, it “measures whether the predicted return direction agrees with causal claims extracted from the article,” and more broadly “measures alignment between model predictions and implied causal statements in financial news” (Sun et al., 30 Sep 2025).

The motivation is tied to regime change. The paper argues that financial text shifts not only semantically but also in its causal structure across macroeconomic periods such as pre-COVID, COVID, post-COVID, and rate-hike regimes. Standard predictive metrics such as MSE can show whether a forecast is numerically accurate on average, but they do not indicate whether the forecast is aligned with the narrative drivers described in the text. FCAS is therefore intended as a diagnostic of whether a model remains aligned with the causal implications of financial narratives when those narratives change.

In this formulation, “alignment with causal cues” is explicitly narrower than structural causality in the econometric sense. FCAS does not test whether a model has recovered true market causation. It tests whether the sign of the model output matches the sign of an extracted causal polarity from the article. The score is thus best understood as a causal-alignment diagnostic based on textual causal polarity, rather than as a causal inference estimator or a feature-attribution method in the XAI sense.

2. Mathematical specification

The paper gives both a population-level functional and a sample estimator. At the regime level, for predictive model $f_\theta$ and regime-specific distribution $\mathcal{P}_{r_k}$, FCAS is defined as

$$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$

Its empirical form is

$$\text{FCAS} = \frac{1}{N} \sum_{i=1}^{N} \mathbb{I}\!\left[\operatorname{sign}(f_\theta(x_i)) = \operatorname{sign}(c_i)\right],$$

where $c_i$ is the polarity of extracted causal cues (Sun et al., 30 Sep 2025).

The variables are specified narrowly. $f_\theta$ is a predictive model mapping financial text to a scalar return prediction; $x_i$ is a financial text input such as a news article; $r_k$ denotes a macroeconomic regime; $\mathcal{P}_{r_k}$ is the regime-specific data distribution; $c_i$ is the extracted causal claim or polarity of extracted causal cues; and $\mathcal{P}_{r_k}$0 is an indicator function. Because the estimator is normalized by $\mathcal{P}_{r_k}$1, the formal sample definition is an average agreement rate over an evaluation set.

The paper is equally notable for what it does not specify. It does not provide an algorithm for extracting $\mathcal{P}_{r_k}$2; it does not state whether $\mathcal{P}_{r_k}$3 is binary, ternary, or real-valued before the sign operation; and it does not define weighting, thresholding, or confidence calibration for the causal-cue extraction step. It also does not introduce any alternative FCAS variants, approximation schemes, attention-alignment versions, gradient-based decompositions, or rationale-overlap formulas. The only formal FCAS definition given is sign agreement.

3. Operational computation and role in the framework

Operationally, FCAS is computed post hoc. The implied procedure is to take a financial news article $\mathcal{P}_{r_k}$4, run the predictive model to obtain $\mathcal{P}_{r_k}$5 as a predicted next-day return, extract from the article a causal claim or causal-cue polarity $\mathcal{P}_{r_k}$6, convert both to directional signs, score the instance as $\mathcal{P}_{r_k}$7 if the signs match and $\mathcal{P}_{r_k}$8 otherwise, and average within a regime (Sun et al., 30 Sep 2025).

This makes FCAS model-agnostic with respect to the prediction architecture. It depends only on the scalar output $\mathcal{P}_{r_k}$9 and an externally extracted causal polarity $$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$0. It does not require access to attention maps, gradients, token-level saliency, hidden-state attribution, or rationale heads. Accordingly, the paper positions FCAS as a post hoc evaluation metric rather than a training objective; no FCAS-based loss term is introduced.

Within the broader robustness profile, FCAS occupies the causal-alignment dimension. The paper summarizes the framework as

$$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$1

FCAS and NLICS are described as capturing alignment with causal and logical content, whereas PCS and TSV capture perturbation sensitivity and semantic drift. This division is conceptually important: FCAS is not a substitute for predictive error, but a complementary diagnostic aimed at exposing structural misalignment between forecast direction and the causal framing of the text.

A practical difficulty follows immediately from this design. Because the paper does not specify a causal lexicon, annotation guidelines, dependency patterns, human labeling protocol, extraction templates, or pseudocode for deriving $$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$2, the central external signal on which FCAS depends remains under-specified. This suggests that any replication necessarily requires additional design choices outside the published metric definition. A plausible implication is that established financial causality extraction tasks and sequence-labeling pipelines—such as those developed in FinCausal 2020 and FinCausal 2021 for cause-effect extraction in financial documents—would be natural candidates for operationalizing the extraction step, but the FCAS paper itself does not make that choice explicit (Mariko et al., 2020, Haldar et al., 2021).

4. Experimental use and reported findings

FCAS is evaluated on FNSPID financial news aligned with next-day returns, with stock prices from Yahoo Finance, over the period 2018–2023. The universe comprises 110 S&P 500 companies across 11 GICS sectors. The four regimes are pre-COVID (2019-11-01 to 2019-12-31), COVID (2020-01-01 to 2020-03-23), post-COVID (2020-05-01 to 2020-07-01), and rate-hike (2022-02-15 to 2022-06-15). Models include an LSTM on TF-IDF, a Text Transformer based on fine-tuned DistilBERT, and a Feature-based Transformer combining TF-IDF and MiniLM embeddings (Sun et al., 30 Sep 2025).

The clearest explicit FCAS values appear in the stock-level case study for JPM and AAPL. For JPM, the reported FCAS values are 1.118 in pre-COVID, -2.096 in COVID, -0.948 in post-COVID, and 2.137 in rate-hike. For AAPL, the corresponding values are 1.146, -2.090, -0.880, and 2.153. The authors interpret this pattern as causal alignment dropping during COVID and rebounding during the rate-hike period.

These case-study values create a major specification problem. Under the formal estimator,

$$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$3

FCAS should lie in $$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$4. The reported values include negatives and values above 1. The paper does not explain this discrepancy. By contrast, the ablation table reports Transformer FCAS values in the rate-hike period of 0.62 for full evaluation, 0.60 with no PCS, 0.58 with no TSV, and 0.61 with no NLICS, which are consistent with the formal definition. The coexistence of these two scales is a serious internal inconsistency.

Beyond those point results, the paper supports only limited FCAS-specific comparative claims. It states that FCAS is used across all regimes and models, that FCAS worsens during COVID in the case study and recovers later, and that FCAS is one of the metrics whose removal produces the largest reductions in interpretability. It does not provide a complete regime-by-model FCAS table, FCAS-specific significance tests, or an explicit correlation analysis between FCAS and MSE. The evidence connecting FCAS to robustness is therefore primarily qualitative rather than fully inferential.

5. Interpretation and relation to adjacent causal notions

Under the formal definition, high FCAS means that predicted return directions usually match the polarity of extracted causal claims, whereas low FCAS means that the model often predicts directions inconsistent with the article’s implied causal drivers. Strong FCAS is associated with predicting upward movement when the article contains positive causal cues, downward movement when it contains negative cues, and maintaining that directional alignment across regimes. Weak FCAS is associated with reliance on spurious lexical shortcuts, regime-specific overfitting, or reversal of the article’s implied causal direction.

The term “causal attribution” in FCAS is deliberately limited. The score does not involve token-level attribution maps, integrated gradients, Shapley values, attention-based explanation, or intervention-based validation. In that respect it differs sharply from several adjacent lines of work. Suppes-Bayes Causal Networks in financial stress testing define prima facie causation through temporal priority and probability raising, and support stress-scenario generation through directed causal structure (Gao et al., 2017). The database-oriented Causal-Effect Score defines attribution as an interventional difference in expectation,

$$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$5

which is substantially closer to a formal intervention-based attribution semantics (Azua et al., 4 Feb 2025). Directed-information methods rank market entities by net predictive information flow and are explicitly described as predictive-information attribution rather than intervention causality (Diamandis et al., 2018). Mechanism-level causal change attribution, in turn, decomposes changes in an outcome distribution across environments using counterfactual mechanism switches and multiply robust estimation (Quintas-Martinez et al., 2024).

Against that background, FCAS is much narrower. It is a sign-based agreement score between output direction and extracted causal polarity. It is therefore closer to textual causal-cue agreement or causal-narrative alignment than to structural causal modeling, counterfactual identification, or formal interventional attribution. This suggests that FCAS occupies a distinct niche: it evaluates whether financial NLP predictions remain behaviorally consistent with narrative causality as expressed in text, not whether the underlying model has recovered economically identified causal mechanisms.

6. Limitations, reproducibility, and practical use

The largest limitation is the under-specified extraction of $$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$6. The paper requires “causal claims extracted from the article,” “implied causal statements in financial news,” and “the polarity of extracted causal cues,” but does not describe how those objects are obtained. As a result, FCAS is not fully reproducible from the paper alone.

A second major limitation is the inconsistent reported scale. The formal definition implies an agreement rate in $$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$7, yet the stock-level case study reports values such as $$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$8 and $$\text{FCAS}(f_\theta, \mathcal{P}_{r_k}) = \mathbb{E}_{(x, c)} \left[ \mathbb{I} \left( \operatorname{sign}(f_\theta(x)) = \operatorname{sign}(c) \right) \right].$$9. Because the paper does not resolve whether this reflects an undocumented variant or a reporting error, FCAS as presented is only partially specified in operational terms.

Additional caveats follow from the metric’s dependence on external causal-cue judgments. Cue selection, polarity assignment, and ambiguity resolution can all inject bias, particularly because financial reporting often contains mixed, implicit, or regime-dependent causal language. The paper’s own empirical narrative—that FCAS degrades in COVID-like periods—shows both the usefulness of the metric and its fragility: the causal language itself may shift most sharply precisely when robustness matters most.

FCAS should therefore not be interpreted as proof that a model has learned true market causation. It is an auditing metric for whether predicted return direction is consistent with extracted textual cause-effect polarity under changing macroeconomic regimes. Its practical value lies in regime-aware diagnostic use: checking whether a deployed financial NLP model remains aligned with narrative drivers when semantic drift, crisis discourse, and causal reframing threaten out-of-distribution reliability.