X-FCA: Diagnosing Explanatory Decoupling in Financial Information Systems via Counterfactual XAI and Conformal Prediction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Automated financial information systems increasingly delegate sentiment-drivendecisions to transformer-based language models, yet whether those decisionsrest on semantic content or on representational shortcuts is invisible to stan-dard accuracy metrics. Analysing a domain-specific financial transformer, weidentify a phenomenon we term explanatory decoupling: a robustness interven-tion can remove a model’s reliance on a spurious cue without altering what themodel attends to, so explanation stability cannot be read as evidence of semanticgrounding. We establish this with X-FCA (eXplainable Financial CounterfactualAnalysis), a diagnostic framework coupling counterfactual perturbation, Con-formal Prediction, and token-level attribution, applied to FinTwitBERT over aharmonized corpus of 177,965 microblogs spanning cryptocurrency, traditionalfinance, and institutional sources, so that any fragility observed reflects thearchitecture rather than a single dataset. Deterministic emoji ablation reversespredictions 2.77 times more often than a matched random-deletion control, pro-viding causal evidence that visual tokens, not sequence-length sensitivity, drivethe fragility. Integrated Gradients trace this to the tokenizer: all ten tested finan-cial emojis, including semantically opposed bullish and bearish pairs, collapse toa single [UNK] token—a pre-inference information loss we term tokenizer blind-ness that leaves the model responding to visual density rather than polarity.Adversarial vaccination removes the coupling between emoji attention and pre-diction confidence (from r=0.121, p=0.007, to r=-0.004, n.s.) and significantlycompresses prediction drift (p < 0.001), yet attributional consistency remainsstatistically unchanged: robustness without semantic regrounding. Conformal prediction converts the residual overconfident errors into calibrated uncertaintysets. Robustness, uncertainty calibration, and attribution analysis are thereforedistinct and non-substitutable axes of reliability assessment: evidence on any onelicenses no conclusion about the others.