Large language models (LLMs) are rapidly becoming integral to financial analysis, from parsing earnings calls to predicting stock market reactions to news. But a critical question remains: When we feed these models more information, do they perform better? Our recent study suggests, not necessarily. We document a structural limitation of LLMs in financial tasks, a phenomenon we call information overload, where too much context leads to worse outcomes.

Drawing on behavioral economics, particularly Herbert Simon’s concept of bounded rationality, we show that LLMs, like humans, can struggle with excess input. However, their limitations are not emotional or cognitive but arise from the technical design of modern neural networks and the way they are trained. These findings raise practical concerns for researchers, analysts, and institutions that increasingly rely on AI for economic forecasting, portfolio construction, and regulatory insight.

The Empirical Pattern: An Inverted U

We test this idea across two empirical domains. First, we use an LLM to forecast earnings changes based on transcripts of corporate earnings calls. We gradually increase the model’s input, from the first few lines of the most recent call to the full transcripts of calls from three prior quarters. Performance initially improves but then declines, a textbook inverted U-shape. The accuracy rises from 52 percent to 55 percent and then drops as more context is added. The F1 score – a measure of predictive performance – follows a similar path, peaking at 65 percent before falling.

Next, we examine this question as to asset pricing. Here, the task is to classify stock-specific news as likely to generate a positive or negative same-day return. Again, we add context incrementally, with up to 40 prior news items. And again, we observe that accuracy initially improves and then deteriorates, eventually falling to near-random levels.

To explore how this affects investment strategies, we construct long-short portfolios using LLM-embedded financial news. When using smaller models, adding contextual information to short alerts does not yield higher returns. However, larger and more advanced models do benefit from additional context, suggesting that only sufficiently capable LLMs can handle more complex inputs without succumbing to informational overload.

The Friction: Why More Can Be Worse

Why do LLMs falter when given more data? We identify two technical frictions.

First, neural networks process information through a layered architecture, with each layer serving as a fixed-size bottleneck. These layers must compress and encode information passed forward, and while they perform remarkably well in many tasks, they are imperfect. As more text is added, the model must filter and prioritize. With limited capacity and flawed encoders, LLMs increasingly waste processing power on irrelevant or redundant content. Unlike a perfectly rational agent, they don’t always know what to ignore.

Second, the way LLMs are trained exacerbates this problem. Most modern LLMs expand how much text they can take in at once, over time. Initial training is focused on short inputs; longer contexts are only introduced in later stages. These models are optimized to find specific “needles in the haystack,” not to systematically ignore unhelpful information. Hence, while an LLM might retrieve a buried fact when prompted directly, it might still falter in holistic tasks requiring synthesis and filtration over long texts.

This design asymmetry means that models often perform better with shorter inputs, even if the extra information is potentially relevant. It also means that empirical testing, not just pure intuition, is needed to find the right context length for a given financial task.

Implications for AI Use in Financial Research

Our findings have several implications for financial professionals, academics, and policymakers.

1. LLMs are not “free lunch” information processors. The belief that more context always yield better performance is flawed. In many cases, models do worse with more data unless the information is carefully curated or the model is sufficiently powerful.
2. Context inclusion is an optimization problem. Analysts and researchers should not treat LLM prompts as “the more, the merrier.” Just as overfitting is a risk in econometric modeling, excessive input can distort LLM outputs.
3. Model size matters. Larger models (e.g., Meta’s LLaMA 3 405B) are more capable of digesting long context effectively. We show that portfolio strategies based on embeddings from these models perform significantly better with additional context than those built using smaller models.
4. Informational biases are not just in training data. Prior work on AI bias in finance has focused on demographic or content biases embedded in the training corpus. Our work highlights a new class of bias rooted in computational limitations. These biases are structural, predictable, and, crucially, independent of training data. They affect how LLMs process real-time inputs, not just what they’ve learned.
5. Regulators and researchers must scrutinize AI “black boxes.” As LLMs become embedded in asset management, credit scoring, and financial compliance, understanding these structural frictions becomes essential. Treating LLMs as rational black boxes may lead to subtle but systematic errors.

Conclusion

Artificial intelligence has the potential to transform financial decision-making, but it is not immune to failure. Our results highlight a key structural limitation: LLMs, like their human counterparts, suffer from information overload. Performance peaks and then declines as more context is added. Larger models mitigate this issue, but none eliminate it.

The implication is clear: Using LLMs effectively in finance requires empirical tuning of inputs, careful model selection, and a shift in mindset. More data is not always better, and black-box AI is no substitute for rigorous methodology.

This post comes to us from Attila Balogh at the University of Melbourne, Antoine Didisheim at the University of Melbourne, Luciano Somoza at ESSEC Business School, and Hanqing Tian at the University of Melbourne. It is based on their recent paper, “AI in Finance and Information Overload,” available here.