[Paper Review] Towards a more efficient bias detection in financial language models

Firas Hadj Kacem, Ahmed Khanfir+1 more|arXiv (Cornell University)|Mar 9, 2026

Stock Market Forecasting Methods0 citations

TL;DR

The paper analyzes bias in five financial language models using a large real-finance sentence dataset, showing shared bias patterns and proposing cross-model guided bias detection to reduce costs.

ABSTRACT

Bias in financial language models constitutes a major obstacle to their adoption in real-world applications. Detecting such bias is challenging, as it requires identifying inputs whose predictions change when varying properties unrelated to the decision, such as demographic attributes. Existing approaches typically rely on exhaustive mutation and pairwise prediction analysis over large corpora, which is effective but computationally expensive-particularly for large language models and can become impractical in continuous retraining and releasing processes. Aiming at reducing this cost, we conduct a large-scale study of bias in five financial language models, examining similarities in their bias tendencies across protected attributes and exploring cross-model-guided bias detection to identify bias-revealing inputs earlier. Our study uses approximately 17k real financial news sentences, mutated to construct over 125k original-mutant pairs. Results show that all models exhibit bias under both atomic (0.58\%-6.05\%) and intersectional (0.75\%-5.97\%) settings. Moreover, we observe consistent patterns in bias-revealing inputs across models, enabling substantial reuse and cost reduction in bias detection. For example, up to 73\% of FinMA's biased behaviours can be uncovered using only 20\% of the input pairs when guided by properties derived from DistilRoBERTa outputs.

Motivation & Objective

• Assess bias in five financial language models (two generative, three encoders) using real financial sentences mutated for protected attributes.
• Explore whether bias-revealing inputs are shared across models and can be reused.
• Evaluate cross-model guided bias detection to reduce computation and inference costs.
• Quantify atomic and intersectional bias across gender, race, and body attributes.
• Identify practical strategies to accelerate bias auditing in financial NLP deployments.

Proposed method

• Mutate real financial sentences using HInter to perform atomic and intersectional attribute changes.
• Run sentiment prediction for original and mutated inputs across five models (FinMA, FinGPT, FinBERT, DeBERTa-v3, DistilRoBERTa).
• Detect bias when original and mutated inputs yield different sentiment labels.
• Compute bias-detection ratios and analyze overlaps of bias-revealing inputs across models.
• Measure prediction shifts with Jensen–Shannon Distance and Cosine Similarity to capture non-flipping biases.
• Evaluate cross-model guided bias detection by prioritizing inputs based on other models’ predictions and compare to random input ordering.

Experimental results

Research questions

• RQ1 Do financial language models exhibit atomic and intersectional bias across protected attributes (gender, race, body)?
• RQ2 Are bias-revealing inputs shared across models, enabling input reuse for bias detection?
• RQ3 Can cross-model guidance (using lightweight models) accelerate bias detection in larger models without sacrificing reliability?

Key findings

• All five models show bias under atomic and intersectional settings with varying magnitudes.
• Lightweight models exhibit lower overall bias than large models in atomic and intersectional settings.
• Overlaps of bias-revealing inputs are high among classifier models (over 94% shared) but low between classifiers and generative models.
• Cross-model bias detection guided by predictions from lightweight models can uncover a large fraction of bias in large models with limited inputs (e.g., 73% of FinMA bias found with 20% of inputs).
• Bias-revealing inputs tend to produce larger prediction shifts (higher JSD) than non-bias-revealing inputs, supporting cross-model prioritization strategies.
• Results support reusing bias inputs across similar model families to reduce auditing costs.