Can LLMs Hit Moving Targets? Tracking Evolving Signals in Corporate Disclosures

TL;DR

Moving targets in corporate disclosures—shifting performance metrics when targets become hard—are linked to stock underperformance, but prior NER-based extraction introduces noise and loses context. This work presents an LLM-driven target extraction pipeline and a semantic-similarity scoring mechanism that replaces NER, enabling semantically rich targets and robust cross-period matching for the Moving Targets MT_c,t. Empirical results using S&P 100 earnings calls show reduced noise, richer targets, and stronger predictive power for subsequent declines than the NER method. The approach enhances financial text-based predictive analytics and yields more interpretable, actionable metrics for evaluating managerial emphasis.

Abstract

Moving targets -- managers' strategic shifting of key performance metrics when the original targets become difficult to achieve -- have been shown to predict subsequent stock underperformance. However, our work reveals that the method employed in that study exhibits two key limitations that hinder the accuracy -- noise in the extracted targets and loss of contextual information -- both of which stem primarily from the use of a named entity recognition (NER). To address these two limitations, we propose an LLM-based target extraction method with a newly defined metric that better captures semantic context. This approach preserves semantic context beyond simple entity recognition and yields consistently higher predictive power than the original approach. Overall, our approach enhances the granularity and accuracy of financial text-based performance prediction.

Figures (5)

• Figure 1: An illustration of the metric computation for LLM-based extraction leveraging semantic similarity. First, we extract targets from the firm’s earnings call transcripts at year-quarter $t$ and at $t-4$ using a LLM. Second, all extracted targets are encoded as vector embeddings using text encoder $f_T(\cdot)$. Third, for each target at $t-4$, we compute pairwise cosine similarities with targets at $t$, apply max-pooling to identify the most similar match, and perform thresholding $g(\cdot)$ to correct for paraphrases and remain robust to wording or formatting variations. Finally, we average the retained similarities to obtain the firm $c$’s Moving Targets score at time $t$, denoted $MT_{c,t}$.
• Figure 2: Comparison of excess returns, Fama–French 3-factor alphas, and 5-factor alphas under the NER-based vs. LLM-based scoring approaches. Estimates are based on a calendar-time portfolio strategy that goes long in firms in the lowest quintile (Q1) of the Moving Targets distribution and short in those in the highest quintile (Q5). This design captures the return spread between firms with relatively fewer abandoned targets and those with more pronounced target shifting. The LLM-based approach consistently delivers stronger predictive power than the NER-based one.
• Figure 3: Representative example of targets that are extracted by NER-based method showing noise in the extracted targets, excerpted from Apple Inc. 2010 Q2 earnings call transcript. The NER-based method extracts generic terms such as year and units as targets illustrating the noisy nature of the extraction.
• Figure 4: Representative example of targets that are extracted by NER-based method showing loss of contextual information, excerpted from NVIDIA Corporation 2025 Q2 earnings call transcript. The NER-based method extracts words like revenue as targets but drops surrounding qualifiers (e.g., Data Center revenue, Compute revenue), losing the specific type of revenue or operational driver. In addition, key product-related terms such as NVIDIA Hopper, and GPU computing are excluded from the target set even though the original approach is designed to include product entities, further underscoring the limitation of the NER-based extraction method.
• Figure 5: Representative example of targets that are extracted by LLM-based method, excerpted from NVIDIA Corporation 2025 Q2 earnings call transcript. Here, we highlight both the phrase in the transcript where each target is detected and the corresponding extracted target itself. Unlike the NER-based approach, which often extracts overly generic terms such as year and units, the LLM-based method captures richer and more semantically precise business metrics (e.g., Quarterly Data Center revenue, Quarterly Compute revenue growth), reducing the noise in the extracted set.