Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Auditing Algorithmic Bias in Transformer-Based Trading

Armin Gerami, Ramani Duraiswami

TL;DR

The paper audits a Transformer-based trading setup by applying Partial Information Decomposition to quantify how much each input stock influences a target's price-mrediction and how this influence relates to implied volatility and trading frequency. It trains a Transformer to predict price movements using a target stock and several supports, then computes Excluded Information to gauge each input's influence. Key findings show the model does not reduce reliance on high-volatility inputs and exhibits a bias toward lower-frequency price movements, with frequency reduction amplifying this bias. The work highlights transparency gaps in financial Transformers and points to frequency-based regularization as a potential mitigating factor, advancing interpretability in algorithmic trading 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.

Auditing Algorithmic Bias in Transformer-Based Trading

TL;DR

The paper audits a Transformer-based trading setup by applying Partial Information Decomposition to quantify how much each input stock influences a target's price-mrediction and how this influence relates to implied volatility and trading frequency. It trains a Transformer to predict price movements using a target stock and several supports, then computes Excluded Information to gauge each input's influence. Key findings show the model does not reduce reliance on high-volatility inputs and exhibits a bias toward lower-frequency price movements, with frequency reduction amplifying this bias. The work highlights transparency gaps in financial Transformers and points to frequency-based regularization as a potential mitigating factor, advancing interpretability in algorithmic trading 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.

Paper Structure

This paper contains 12 sections, 3 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Background
  3. Transformer Models
  4. Partial Information Decomposition
  5. Method
  6. Training Transformer
  7. Measuring Influence
  8. Results
  9. Reliance Independent of IV
  10. Trade Frequency Bias
  11. Conclusion
  12. Supporting Figures

Figures (3)

  • Figure 1: Excluded Information (EI) for each support stock used when predicting the target stock, NVDA. A higher EI value indicates that the model relies less on that stock's data. The right axis displays the 30-day implied volatility (IV) as a proxy for risk. Ideally, high IV should lead to high EI (less reliance on risky data), but the near-zero correlation shown here indicates the model is not discounting this risk. The plot also includes scenarios where support stocks are modified to simulate 3× and 5× reduced trading frequencies for the support stocks, which consistently results in lower EI.
  • Figure 2: Percentage return and 30-day implied volatility of our target stock (NVDA), and the support stocks (AMD, TSM, MU, INTC).
  • Figure 3: The actual percentage return of the NVDA stock and the trained transformer's prediction.