Uncovering Representation Bias for Investment Decisions in Open-Source Large Language Models

TL;DR

This work investigates representation bias in financial open-source LLMs (Qwen family) by applying a balanced round-robin, category-driven prompting framework to roughly 150 US equities. It derives firm-level confidence scores from token-log-probabilities and aggregates them to quantify model preferences, linking confidence to firm features via multiple statistical tests. The findings show that firm size and valuation strongly shape LLM confidence, with pervasive cross-context anchoring and sector-specific patterns; category prompts ground preferences toward fundamental metrics like free cash flow, with weaker ties to growth and higher risk. The study highlights the need for sector-aware calibration and robust evaluation protocols to ensure safe, fair deployment of financial LLMs, and outlines future debiasing and mechanistic analyses.

Abstract

Large Language Models are increasingly adopted in financial applications to support investment workflows. However, prior studies have seldom examined how these models reflect biases related to firm size, sector, or financial characteristics, which can significantly impact decision-making. This paper addresses this gap by focusing on representation bias in open-source Qwen models. We propose a balanced round-robin prompting method over approximately 150 U.S. equities, applying constrained decoding and token-logit aggregation to derive firm-level confidence scores across financial contexts. Using statistical tests and variance analysis, we find that firm size and valuation consistently increase model confidence, while risk factors tend to decrease it. Confidence varies significantly across sectors, with the Technology sector showing the greatest variability. When models are prompted for specific financial categories, their confidence rankings best align with fundamental data, moderately with technical signals, and least with growth indicators. These results highlight representation bias in Qwen models and motivate sector-aware calibration and category-conditioned evaluation protocols for safe and fair financial LLM deployment.

Figures (2)

• Figure 1: Determinants of LLM confidence by model and correlation type.
• Figure 2: (Left) Effect sizes across models and factors. (Middle, Right) Heatmap of cross-context stability across sectors and models.