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

Summary

• The X-FCA framework was established to audit financial language models over a harmonized master corpus of 177,965 microblogs across Cryptocurrency, Traditional Finance, and Institutional sources.p.2
• Integrated Gradients revealed tokenizer blindness, showing that all ten tested financial emojis deterministically collapse into a single [UNK] token with ID=2.p.11

Key insights

1. Aggregated validation metrics like Accuracy and F1 conceal fundamental operational model brittleness under non-alphanumeric perturbations.p.5
2. Tokenizer blindness causes all financial emojis to collapse to [UNK] prior to inference, leaving models to respond to visual token density rather than semantic polarity.p.4
3. Model fragility follows a non-linear density-dependent Threshold Effect, spiking to 10.0% instability at clusters of three or more emojis.p.4
4. Robustness interventions can remove spurious reliance on visual cues without altering token-level attributions, demonstrating explanatory decoupling.p.4
5. Adversarial vaccination completely eliminates spurious coupling, shifting the correlation between emoji attention and prediction confidence from r=0.121 (p=0.007) to r=-0.004 (n.s.).p.20