Caution Ahead: Numerical Reasoning and Look-Ahead Bias in AI Models

Abstract

Recent work within accounting and finance has highlighted that modern AI systems exhibit superhuman performance on a variety of foundational activities within these fields. However, the literature often does not provide economic rationale forwhyAI models seem to outperform, largely because these models are a black box. Through a series of experiments, I set out to open the black box and provide direct evidence onhowandwhyAI models appear to perform so well on accounting and finance-related tasks. I show that much of the superior performance of AI models can be attributed to artifacts of the modeling itself, rather than to mechanisms grounded in economics. Focusing on two key components of AI models, which may bias inferences in papers that rely on them, I first show that Large Language Model (LLMs) exhibit extremely poor numerical reasoning and thus application in these settings should proceed with caution. Second, I highlight that commercial LLMs suffer from significant look-ahead bias, which may explain a large portion of their predictive ability in various settings.