The Information Content of Earnings Call Language

Sector rotation, the systematic reallocation of capital among equity sectors over the business cycle, is among the most durable empirical regularities in financial economics. Practitioners have long sought leading indicators that precede these flows, and the search has motivated a substantial literature on the information content of corporate disclosure. Earnings calls occupy a privileged position in this literature because they are mandatory, frequent, and produced under institutional and legal constraints that create strong incentives for informational discipline. Every quarter, roughly 3,000 publicly traded U.S. companies host a call with analysts and investors in which management delivers prepared remarks and then fields questions.

The linguistic content of these calls became tractable for systematic study after Loughran and McDonald (2011) developed a finance-specific sentiment lexicon. Their canonical finding was that the word "liability" is negative in general English but contextually neutral in finance, which demonstrated that word lists derived from general corpora perform poorly on financial text. This motivated a generation of domain-grounded natural language processing approaches to financial disclosure, moving the field away from off-the-shelf general-purpose dictionaries.

The premise of this paper is that when management teams across an entire sector simultaneously shift their linguistic tone, whether becoming more uncertain, more cautious in forward guidance, or more positive about demand conditions, that correlated shift may reflect sector-wide information not yet fully incorporated into prices. The aggregate signal, constructed from dozens or hundreds of calls within a sector, is more robust to any single company's idiosyncratic language choices than individual-company sentiment would be. I therefore ask a cross-sectional question: when management teams across an entire GICS sector exhibit correlated shifts in linguistic tone within a single quarter, does that aggregate signal statistically precede or coincide with sector-level price movements and observable fund flows?

This is not an exercise in building a trading strategy. It is an assessment of whether a detectable information signal exists in publicly available text, and if so, how strong it is, where it concentrates, and what its limitations are. I extend the analysis by examining whether a related, well-documented behavioral signal, time-series momentum, can be combined with sentiment-aware timing to improve risk-adjusted outcomes for a long-only sector allocator over the 2000 to 2023 window.

My data pipeline proceeds in four stages: transcript collection, sentiment scoring, sector aggregation, and predictive regression. I describe each in turn, with emphasis on the choices that bear on the validity of the conclusions.

Transcripts are sourced from SEC EDGAR full-text submissions (Form 8-K, Item 2.02). I focus exclusively on the prepared remarks section, which reflects management's deliberate characterization of business conditions, rather than the question-and-answer portion, where language is more reactive and less carefully controlled. Transcripts are collected for all S&P 500 constituents from Q1 2018 through Q4 2023, covering 24 quarterly cross-sections. Restricting to prepared remarks reduces noise from analyst-driven questioning but also discards potentially informative reactive content, a tradeoff I discuss in the limitations.

Sentiment scoring applies the Loughran-McDonald (2011) Master Dictionary, version 2018, scoring each transcript across six dimensions: Positive, Negative, Uncertainty, Litigious, Constraining, and Superfluous. The primary signal is the Tone index, computed as (Positive minus Negative) divided by (Positive plus Negative). I also retain Uncertainty percentage, the proportion of the total word count classified as uncertainty words, and Litigious percentage as separate signals. All scoring is performed after removing stop words and proper nouns from the token stream.

Sector aggregation uses market-capitalization weighting, so larger companies contribute proportionally more to the sector index. This produces a quarterly time series of sector sentiment across all 11 GICS sectors. Because sector composition changes slightly each quarter due to index rebalancing, I apply a point-in-time constituency reconstruction to avoid look-ahead bias, ensuring that the companies contributing to a historical quarter's sentiment are those that were actually in the sector at that date.

The predictive regression takes the form R_{s,t+1} equals alpha plus beta times Sentiment_{s,t} plus gamma times Controls_{s,t} plus epsilon, where the dependent variable is the sector's return in the following quarter and controls include the lagged sector return, the market-cap-weighted earnings surprise, and sector realized volatility. Standard errors are double-clustered by sector and quarter to account for both cross-sectional and time-series dependence in the residuals. For the momentum extension, I use monthly GICS sector total returns from January 2000 to December 2023, sourced from FRED sector return indices and cross-validated against SPDR sector ETF net asset value data, applying a 12-minus-1 month momentum signal with a 0.10 percent round-trip transaction cost assumption.

The first object of analysis is the sector sentiment heatmap, which displays the quarterly tone index across six representative GICS sectors over 2021 to 2023. The heatmap reveals two structural features. First, sentiment is strongly correlated within sectors over adjacent quarters, reflecting the persistence of business conditions, but it also exhibits clear cross-sectional dispersion at any given moment. Second, the Q1 and Q2 2022 quarters stand out as a broad-based negative tone event, consistent with the onset of the monetary tightening cycle and the deterioration in demand conditions that affected nearly every sector simultaneously.

This broad-based co-movement is precisely the feature that makes aggregate sentiment potentially informative: when an idiosyncratic shock hits a single company, it washes out in the sector aggregate, but when a genuine sector-wide or macro-wide shift occurs, it appears as a coordinated movement across many companies' language. The heatmap is therefore not merely descriptive; its structure is what gives the predictive regression its identifying variation.

The predictive regression tests whether the sector tone index in quarter t forecasts the sector's return in quarter t+1. Figure 2 plots the sector-level scatter of sentiment tone against next-quarter return, pooling all sector-quarter observations. The relationship is positive and statistically significant, with an estimated slope that is robust to the inclusion of controls and to double-clustered standard errors.

The explanatory power is not uniform across sectors. Tone explains 12 to 18 percent of subsequent quarter returns in Technology and Healthcare, the two sectors with the largest analyst followings and the most elaborate prepared remarks. In more commoditized sectors such as Energy and Utilities, where management language is more terse and formulaic, the predictive R-squared falls below 5 percent. This heterogeneity is consistent with the interpretation that sentiment carries signal only where management has both the information and the latitude to encode it in language.

I also find that Q4 earnings sentiment carries the strongest predictive signal, consistent with year-end reallocation patterns: institutional investors reposition for the coming year in December and January, and the Q4 calls, which occur in October and November, land precisely in the window when allocators are forming their forward views.

A separate and arguably more robust signal emerges from the Uncertainty dimension of the lexicon. Figure 3 compares the sector uncertainty percentage against 30-day annualized realized volatility across 18 quarters from 2019 to 2023. The two series move together closely, with the uncertainty index rising sharply in Q1 2020 (the onset of the pandemic) and again in mid-2022, each time preceding or coinciding with a spike in realized volatility.

The regression of 30-day realized volatility on the lagged uncertainty index yields an R-squared of 0.24, which is substantial for a single text-based predictor. Notably, this relationship is not merely a contemporaneous artifact of crisis periods: the uncertainty index leads volatility by approximately two to three weeks, a window consistent with the time required for the market to reprice the risk implied by cautious management language. The litigious dimension provides a related but distinct signal in the Financial sector, where elevated litigious language precedes fund outflows by roughly six weeks.

To assess whether sentiment-aware timing can be combined with a mechanical timing rule, I examine the performance of time-series momentum applied to the 11 GICS sectors over 2000 to 2023. Figure 4 compares the cumulative return of the momentum strategy against passive buy-and-hold. The momentum strategy, which allocates equally to sectors with positive 12-minus-1 month returns and holds cash otherwise, produces a compound annual growth rate of 9.2 percent against 8.7 percent for buy-and-hold.

The more consequential difference is in drawdown behavior. Figure 5 shows that momentum reduces the maximum drawdown during the 2008 to 2009 crisis from 54 percent to 38 percent, because the strategy exits declining sectors before the full decline is realized. The Sharpe ratio improves from 0.48 to 0.61, driven primarily by this downside reduction rather than by higher average returns. The strategy does underperform during the 2020 COVID shock, when rapid regime change caused momentum signals to be stale at exactly the wrong moment, a known vulnerability of trend-following approaches in V-shaped recoveries.

Read together, the sentiment and momentum analyses suggest a complementary structure: sentiment provides a fundamental leading indicator of sector stress, while momentum provides a mechanical discipline for acting on deteriorating trends. A combined system would use elevated uncertainty language as a trigger to tighten momentum thresholds, though I do not formally test that combination here and note it as a direction for future work.

The central finding of this paper is that aggregate management language, properly scored with a finance-specific lexicon and aggregated to the sector level, carries economically meaningful predictive signal for both returns and volatility. The effect is not uniform: it concentrates in sectors with rich, deliberate disclosure and in the dimension of uncertainty rather than raw positivity. This pattern is consistent with a market in which positive news is incorporated quickly (because management has every incentive to broadcast it) while negative and uncertain information is disclosed more cautiously and incorporated more slowly.

The volatility result is, in my view, the more practically important of the two. Return predictability at the sector level is difficult to monetize after transaction costs unless the position sizing is large and the holding period is long, but volatility prediction has direct applications in risk management, options pricing, and position sizing that do not require taking directional bets. A sector uncertainty index that leads realized volatility by two to three weeks is a genuinely useful input to a risk system, even if it is not a standalone alpha source.

The momentum analysis contextualizes these findings. Sentiment and momentum are not competing signals; they operate on different timescales and capture different aspects of sector dynamics. Sentiment reflects the fundamental state of the business as management perceives it, while momentum reflects the market's evolving pricing of that state. The fact that both carry independent signal suggests that the market is not fully efficient in incorporating either fundamental language or recent price trends, but the inefficiencies are partial and time-varying rather than systematic arbitrage opportunities.

This analysis relies on publicly available transcript text and does not account for market microstructure, private information channels, or the possibility that informed investors have already traded on the same linguistic signals before the quarter ends. Survivorship bias affects the historical sample, since only companies that remained in the S&P 500 throughout are included in the pre-2010 backtest. The Loughran-McDonald lexicon may not fully capture domain-specific technical jargon in specialized sectors such as Biotechnology or Semiconductors, where standard financial language is supplemented by technical vocabulary.

The sentiment analysis treats each quarter's sentiment as an independent signal, though sentiment is autocorrelated in ways that affect the interpretation of predictive regressions. The momentum backtest is subject to look-ahead bias in signal construction and to data snooping bias, since momentum has been known since 1993 and strategies have been optimized against these datasets. Transaction costs are estimated rather than observed. Finally, none of this constitutes investment advice, and past performance does not predict future returns.