BiasExplainabilityAccountabilitycs.LG
Auditing Algorithmic Bias in Transformer-Based Trading
Armin Gerami, Ramani Duraiswami
October 1, 2025
arXiv.org
Key Takeaways
-
Transformer models ignore market volatility: The study found that transformer-based trading models completely disregard price volatility when making predictions, which is concerning since volatility is a fundamental risk indicator that human traders and traditional models typically consider carefully.
-
Models show bias toward stable, slow-moving assets: The research revealed that these AI systems are biased toward assets with lower-frequency price movements, meaning they may overestimate their ability to predict and trade less volatile stocks while potentially underestimating risks in more volatile markets.
-
Popular AI tools in finance lack adequate bias auditing: The paper highlights that transformer models have become widely used in financial applications despite their potential risks and biases remaining "under-explored," suggesting the financial industry may be deploying these systems without sufficient safety checks.
-
New measurement tools can expose hidden model biases: The researchers introduced a metric based on Partial Information Decomposition (PID) that can measure how much individual assets influence a model's trading decisions, providing a concrete method for auditing and identifying problematic patterns in AI financial systems.
Abstract
Transformer models have become increasingly popular in financial applications, yet their potential risk making and biases remain under-explored. The purpose of this work is to audit the reliance of the model on volatile data for decision-making, and quantify how the frequency of price movements affects the model's prediction confidence. We employ a transformer model for prediction, and introduce a metric based on Partial Information Decomposition (PID) to measure the influence of each asset on the model's decision making. Our analysis reveals two key observations: first, the model disregards data volatility entirely, and second, it is biased toward data with lower-frequency price movements.