1 Introduction

Despite its central role in firm decisions and macroeconomic outcomes—e.g., inflation (Phillips, 1958)—, measuring marginal labor costs remains challenging. While several aggregate indicators (e.g., unemployment and output gap) are commonly used to gauge labor cost pressures, there is little consensus on the most appropriate measure of slack (Barnichon and Shapiro, 2024) and no single measure fully captures true marginal labor costs (Gal and Gertler, 1999). Moreover, their aggregate nature precludes any analysis of firm-level heterogeneity, thereby masking the granular shocks that drive price-setting behavior.

We address these limitations by developing a novel measure of marginal labor cost pressures from textual analysis of earnings calls combined with financial data. These quarterly calls contain rich firm-level information about traditionally elusive variables, e.g., political risks (Hassan et al., 2019) and cost of capital (Gormsen and Huber, 2024). We analyze 248,437 transcripts by US headquartered firms spanning 2002–2025 and find that 86% discuss labor-related issues, revealing granular, real-time information about labor market conditions that aggregate data miss.¹ Building on this rich data, our paper makes two contributions. First, we develop a theoretically grounded approach to translate qualitative discussions into quantitative measures of marginal labor costs. Second, we use this measure to examine how labor cost pressures drive inflation at both aggregate and industry levels, and how individual firms respond to rising labor costs.

We begin by classifying labor-related excerpts into topics employing two approaches: (i) supervised, where we manually classify excerpts into interpretable, pre-specified topics, (ii) unsupervised, where latent topics represent principal components of high-dimensional representations (embeddings) of these excerpts.² For our baseline, we follow the supervised approach. Through manual reading of transcripts, we identify five primary labor-related topics: 1) labor costs, 2) labor shortages, 3) headcount, 4) labor agreements, and 5) labor efficiency. Executives often qualify the discussions of costs, headcount, and efficiency by direction—higher or lower—, leaving us with \(J=8\) topics. To construct dictionaries, we begin with seed keywords directly indicative of each topic such as “labor shortage”, “labor costs”, and “wage inflation” and expand them using embeddings trained on earnings conferences calls following Hassan et al. (2024a). To measure firm \(i\)’s exposure to topic \(j\) in quarter \(t\), \(\Lambda _{it}^{j}\), we count sentences containing these keywords and normalize it by call length.

While our textual analysis provides rich descriptive evidence of labor market pressures, a more structured approach is required to quantify the economic importance of this qualitative data. We develop this framework through the firm’s cost minimization problem that features two variable factors, intermediate inputs and labor. Motivated by research emphasizing the role of non-wage amenities and frictions in labor markets (Rosen, 1986; Hwang et al., 1998; Hall and Mueller, 2018; Bagga et al., 2025), we recognize that firms facing labor cost pressures incur generalized labor costs comprising both direct wage costs and indirect expenses, such as job posting, hiring, training, and retention costs as when expanding their effective workforce.

In our framework, cost-minimizing firms equate marginal revenue products per dollar across variable inputs: An increase in marginal labor cost—holding worker productivity and output elasticities constant—should increase the share of spending on intermediate inputs.³ We therefore estimate each topic’s contribution on changes in marginal labor costs by regressing changes in firms’ revenue shares of intermediate inputs on the intensity of labor discussions of our topics—\(\Lambda _{it}=[\Lambda _{it}^{1},\Lambda _{it}^{2},...,\Lambda _{it}^{J}]\)—while controlling for revenue per employee growth, and firm and time fixed effects, thereby exploiting the within-firm variation. This approach translates our high-frequency qualitative multidimensional discussions from earnings calls into a single quantitative measure of labor cost pressures (\(\omega _{it}\)). The regression coefficients determine the optimal weights for combining different labor topics into a unified measure of their impact on firm labor costs. Finally, the theoretical structure also prevents overfitting—a common pitfall when working with high-dimensional textual data (Gentzkow et al., 2019).

We first validate this measure by comparing a sales-weighted average of \(\omega _{it}\)—an aggregate index of labor cost pressures (\(\omega _t\)) developed from our measure—with traditional aggregate indicators. Our index exhibits strong correlations: 0.75 with labor market tightness (V/U), –0.46 with unemployment, and 0.64 with the Employment Cost Index (ECI). It reveals an intriguing nonlinear relationship, sharply increasing when unemployment falls below roughly 5% and labor market tightness exceeds around 1.5. Notably, following a sharp but short-lived collapse during the Great Recession, our index remained persistently low throughout the subsequent recovery, consistent with the flattening of the Phillips curve (Powell, 2018). However, it surged dramatically during and after the COVID-19 pandemic, aligning with the steepening Phillips curve (Domash and Summers, 2022). When aggregated at the industry level, our measure maintains a robust correlation with labor market tightness and earnings growth of new hires, further validating its relevance. Despite strong correlations with aggregate and industry-level slack variables, these observables explain very little of the variation in firm-level labor cost pressures, suggesting our measure captures idiosyncratic shocks such as localized wage competition and firm-specific expansion plans.

We next quantify the pass-through from labor cost pressures to inflation using industry-level variation in the Producer Price Index (PPI). We find a strong and statistically significant relationship between increased industry-level labor cost pressures and subsequent PPI inflation even after controlling for industry and time fixed effects. Specifically, our estimates indicate that a 1.0 percentage point (pp) rise in labor cost pressures corresponds to approximately a 0.4 to 0.6 pp increase in PPI inflation over the following year. Importantly, the strength of this pass-through varies substantially across industries: sectors with higher labor intensity, such as wholesale trade, retail trade, and accommodation and food services, demonstrate stronger pass-through effects, whereas manufacturing exhibits a notably lower sensitivity along with heavily regulated industries, like utilities and healthcare.

Further, we evaluate the predictive power of the aggregate labor cost pressure measure for core Personal Consumption Expenditures (PCE) inflation using a Phillips curve framework, comparing its performance to traditional measures of labor market slack á la Barnichon and Shapiro (2024). We aggregate our quarterly industry-level measure using PCE weights and find that our measure outperforms traditional labor market indicators—such as unemployment rates, labor market tightness, the output gap, and the ECI—in explaining inflation dynamics. Our measure can explain 19.5 pp of the 29.3 pp cumulative surge in inflation during the post-pandemic period (2021–2022). Our measure also outperforms a “Raw Text” benchmark that includes the eight topic frequencies directly without the theoretical weights, confirming that the structure imposed by the firm’s first-order condition is essential.

When included alongside these traditional measures, our indicator remains statistically significant and subsumes their explanatory power. To investigate the sources of this predictive power, we decompose our measure at the firm-level into a linear “systematic” component (comoving with aggregate slack) and an “idiosyncratic” residual (orthogonal to aggregate slack). We find that the idiosyncratic residual robustly predicts inflation in Phillips-curve regressions, whereas the systematic component does not. This result highlights the critical role of our measure in capturing firm-level convexities and granular shocks in labor cost pressures. For instance, when the marginal cost curve is non-linear, the dispersion of firm-level constraints generates net inflationary pressure that aggregate slack variables miss.

Finally, we exploit the disaggregated nature of our measure to examine how firms respond to labor cost pressures beyond price adjustments. We analyze responses in investment rates, R&D spending, and productivity measures. Our empirical strategy employs a comprehensive set of controls including firm and year fixed effects and industry and firm characteristics to isolate the relationship between labor cost pressures and firm investment behavior.

Firms experiencing higher labor cost pressures increase their investment, with this effect particularly pronounced in industries that heavily employ routine manual workers. These differential investment responses translate into productivity outcomes as well. Firms facing labor cost pressures in routine-manual task-intensive industries experience faster productivity growth, a one standard deviation increase corresponds to a 1.55 percentage point increase in productivity growth. These findings suggest that labor cost pressures accelerate automation and technological adoption, particularly in industries where human labor can be more readily substituted with capital. They also explain why the pass-through of labor cost pressures to PPI inflation is highest for services and near-zero for manufacturing.

Related Literature. Our paper contributes to two strands of literature using textual data in economics. The first employs textual analysis to study previously unmeasurable phenomena such as sentiment and preferences embedded in FOMC transcripts and press conferences (Gorodnichenko et al., 2023; Ahn et al., 2025b; Howes et al., 2026; Ahn et al., 2025a), economic and political uncertainty (Baker et al., 2016; Gorodnichenko et al., 2021; Hassan et al., 2019; Hassan et al., 2024b; Hassan et al., 2023), attention to macroeconomic variables or financial constraints (Song and Stern, 2024; Buehlmaier and Whited, 2018), labor shortages (Harford et al., 2024), and non-wage clauses in firm-worker negotiations (Arold et al., 2024). We extend this method by quantifying firm-level textual measures using a theoretical framework.⁴ This proves essential: raw counts of labor discussions—the prior approach—show no correlation with aggregate slack variables or inflation, whereas our theoretically-grounded measure exhibits strong correlations with both, demonstrating that economic theory substantially improves the predictive power of textual data.

Second, a recent literature uses natural language processing to forecast macroeconomic variables. Most studies time-aggregate their text signal—e.g., averaging daily news-sentiment or collapsing a quarterly central-bank press-conference transcript into a single embedding—so that both left- and right-hand variables share the same frequency (e.g. Ashwin et al. (2024); Araujo et al. (2025); Gosselin and Taskin (2024)). With only 140 quarterly or 300 monthly post-1990 observations, this top-down approach faces severe degrees-of-freedom constraints. We instead discipline the text at the firm level, leveraging around 250,000 firm-quarters. Guided by a cost-minimization framework, we map each firm’s labor-related discussion into a marginal labor-cost pressure and then aggregate these using theory-based weights. This design treats economic structure as a regularizer for high-dimensional text.

Our paper also contributes to a large literature using the Phillips curve to forecast inflation. First, while existing work relies almost exclusively on aggregate slack indicators—such as the unemployment rate (Stock and Watson, 1999), the output gap (Gal 2015), the job-switching rate (Moscarini and Postel-Vinay, 2012), and labor market tightness (Barnichon and Shapiro, 2024)—we construct a novel firm-level measure of labor cost pressures from earnings call transcripts. Our measure directly captures executives’ concerns about labor market conditions, avoiding assumptions about which aggregate variables reflect these conditions or whether they affect all firms uniformly. When aggregated using PCE weights, our measure outperforms all standard indicators in Phillips curve estimation.

Second, exploiting firm-level granularity, we show that small changes in unemployment or output gap rarely generate widespread labor cost discussions. However, during extremely tight labor markets, these concerns become widespread, translating into higher labor cost pressures and inflation. Our analysis shows that firm level labor cost pressures are granular and this granularity should also enter Phillips curve design (e.g., Gabaix (2011)).

This paper also relates to the literature on capital–labor substitution and productivity. Studies such as Acemoglu and Restrepo (2022b) and Graetz and Michaels (2018) show that automation can reduce firms’ reliance on labor. We show that labor cost pressures prompt capital-deepening investments—particularly in industries with routine manual work—often spurring productivity gains. Our findings thus reinforce automation as a critical margin through which firms adapt to labor market shocks (e.g., Leduc and Liu (2024)), with implications for aggregate productivity and optimal monetary policy.

The rest of the paper is organized as follows: Section 2 presents the earnings call data as well as the details of the textual methodology. In Section 3, we introduce the simple theoretical framework used in our estimation. Sections 4 and 5 present details of labor cost pressure estimates and the inflation pass-through. In Section 6 we present how firms respond to changes in labor costs. Finally, Section 7 concludes.

2 Earnings Call Data and Empirical Methodology

Our main source of data is earnings conference calls, where executives discuss results with analysts and investors. We use 248,437 transcripts of earnings conference calls by US headquartered firms held between 2002 and 2025. We begin by constructing a dictionary of 104 labor-related terms (such as ‘personnel’, ‘wage’, and ‘workforce’). To identify these terms, we follow the methodology in Hassan et al. (2024a) and begin with seed keywords (e.g., ’personnel’, ’wage’, ’workforce’) and expand using word embeddings.⁵ For each candidate term, we manually validate that at least 7 of 10 randomly sampled excerpts genuinely discuss labor issues. About 86% of earnings calls in our sample contain at least one of the 104 labor-related terms. Table A.1 shows the firms with the highest average labor discussion intensity throughout our sample period. Notably, staffing and professional services firms dominate the list, reflecting that labor costs and availability are central to their business operations (see Appendix 7 for details).

Table 1 shows sample excerpts from earnings calls containing labor-related discussions. These transcripts reveal several important patterns. First, executives discuss not only labor cost increases but also labor shortages and tight labor market conditions. Second, labor cost pressures extend beyond wages to include sign-on bonuses, retention costs, and other hiring expenses—highlighting that executives conceptualize labor costs more broadly than standard wage measures capture. Finally, executives frequently emphasize the unexpected nature of these developments, describing how labor market conditions “hit us faster than we anticipated” or produced “higher than anticipated wage pressure”. These excerpts demonstrate that earnings calls provide rich, real-time information about firms’ labor cost pressures that may not be fully reflected in traditional aggregate measures.

Figure 1 shows the percentage of earnings calls that contain labor-related discussions over time from 2002 to 2024. While more than 80% of calls contain labor discussions throughout our sample period, the time series shows distinct patterns across different economic periods. During the pre-financial crisis period (2002–2007) around 85–90% of earnings calls consistently contain labor-related discussions. After the 2008–09 financial crisis the percentage drops to around 85 percent and remains at this lower level through most of the 2010s. The data show another shift during the COVID-19 pandemic, with labor-related discussions spiking to almost 100% in 2021 reflecting the unusual labor market disruptions during this period. After this spike the percentage gradually declines but remains elevated and volatile which suggests ongoing labor market adjustments in the post-pandemic period.

Figure 1: Earnings calls with labor mentions over time, %

While these excerpts and their patterns signal the attention executives pay to the labor markets, they do not directly provide a quantitative measure of labor cost pressures. Our next step is to organize these discussions into structured topics and to quantify each topic’s importance for labor cost pressures separately. We employ two complementary approaches: supervised and unsupervised.

In the supervised approach we use our judgment and manual reading of excerpts to define a small number of interpretable labor topics, such as higher labor costs or labor shortages. We then build dictionaries that map sentences into these topics, following the procedure in Hassan et al. (2024a). This supervised strategy produces transparent and easily interpretable topics with clear links to economic concepts.

In the unsupervised approach we let the data speak more freely. We represent each labor-related sentence using pre-trained sentence embeddings and apply principal component analysis to the resulting high dimensional vectors. This yields a set of latent labor topics that are estimated purely from textual patterns. We then replicate our main analysis using these unsupervised topics. Comparing results across both approaches confirms that our findings are not driven by particular supervised labels and show that the information we extract is also present in more flexible data-driven representations of labor discussions.

2.1 Supervised Approach

From reading a random sample of transcripts, we identify five core topics: labor costs, labor shortages, headcount, labor efficiency and labor agreements. When discussing labor costs, headcount and efficiency, the executives also qualify the direction of shift—higher or lower—yielding eight topics total.

To label these topics automatically, we develop dictionaries for each topic following Hassan et al. (2024a). We begin with a small set of seed keywords that directly relate to each topic. For instance, for the topic of labor shortages, we started with words like labor and shortage. Similarly, for labor costs, we considered terms such as wage and compensation. To expand our initial keyword list, we employ word embeddings trained on the earnings call transcripts Pennington et al. (2014). This helped us identify additional words and phrases used in similar contexts. For example, the embedding model suggested terms like personnel shortage and staffing constraints as contextually similar to labor shortage.

In parallel, we manually review excerpts to identify additional frequently used terms missed by the embedding model. To ensure relevance, we extract 10 excerpts containing each candidate keyword from our dataset and manually assess whether they aligned with the intended topic. We retain a keyword in the dictionary for a topic if at least 70% of the sampled excerpts are correctly classified under that topic (i.e., true positives). If a keyword is found to be frequently ambiguous or misclassified, it is either discarded or reassigned to a different topic where it is more relevant. We repeat this process iteratively, adding to the keyword lists with each iteration until no additional keywords emerge. Table A.2 shows the top 20 keyword combinations for each topic from a total of 13,566 keyword combinations across all topics.

Our classification model operates at the sentence level following Hassan et al. (2024a). A sentence is assigned to a topic if it contains all the constituent words from one of that topic’s keyword combinations. For instance, a sentence is categorized under ‘labor shortage’ if it includes both the keywords ‘labor’ and ‘shortage’, regardless of the order or distance between them. We then aggregate across the earnings call transcript, defining firm \(i\)’s exposure to topic \(j\) of labor-related issues in quarter \(t\) \(\Lambda _{it}^{j}\), as the percentage of sentence in an earnings call transcript which mention a topic \(j\).

Figure 2 shows the incidence of these topics in earnings conference calls. These topics have significant overlaps as executives discuss multiple topics in a call. At least one of these topics is mentioned in 62% of earnings conference calls, accounting for 70.4% of overall executive discussions of labor-related terms. The four most prominent topics are higher and lower labor costs, and higher and lower head counts (hiring and firing). These four topics themselves are mentioned in 60% of earnings calls.

Figure 2: Percentage of Earnings Calls, by Topic

Notes: The figures show percentage of earnings calls by labor topic between 2002 and 2025.

2.1.1 Validation of Labor Topic Discussions

We validate that these topics capture genuine labor market conditions by showing that time-series patterns align with business cycles and industry-level topic intensity predicts wage growth for new hires.

Figure A.1 shows each topic evolves cyclically from 2002 to 2025. Discussions of higher labor costs and hiring peaked before the Great Recession, while mentions of lower costs and layoffs surged during downturns. Most striking is the post-pandemic period, when labor shortage discussions reached levels nearly four times higher than any previous peak.

Figure 3 confirms these topics correlate with standard indicators as expected: labor shortages, higher costs, and hiring correlate positively with labor market tightness (V/U ratio), output gap, and the Employment Cost Index, and negatively with unemployment. Discussions of lower costs and layoffs show the opposite pattern.

Figure 3: Correlation of Labor Topic Discussions with Aggregate Variables

Notes:

The table shows piecewise correlations of tightness (calculated as the ratio of postings to vacancies),

unemployment rate, employment cost index, and output gap with labor topics. Each coefficient is calculated

using a regression of one labor market indicator (standardized) on one labor topic (standardized).

Observations are quarterly. To construct labor topic observations at the quarter level we take sales weighted

averages across firms who hold earnings calls in a quarter. Standard errors are robust.

We also investigate how labor-related discussions in earnings calls relate to wage growth for new hires at the industry level. Table A.3 shows that industry-level discussions of higher labor costs strongly predict wage growth for new hires (coefficients 0.62 to 1.58), while mentions of lower costs show negative associations (–0.47 to –1.39). These relationships survive industry and time fixed effects, demonstrating our measures capture meaningful variation in labor market tightness beyond what aggregate time trends or industry composition explain.

3 Theoretical Framework

While our textual analysis of earnings calls provides rich descriptive evidence of labor market pressures, we need a more structured approach to quantify their economic importance on a firm’s marginal costs. In this section we develop a framework through the firm’s profit maximization problem with generalized labor costs, which capture both direct wage costs and indirect labor expenses, such as job posting, hiring, training, and retention costs. This framework recognizes that firms facing labor market pressures incur multiple types of costs as they attempt to expand their effective workforce as we have seen in the discussions of labor topics in the previous section.

In our framework, firm \(i\) in quarter \(t\) simultaneously chooses between its quantity of intermediate inputs, \(M_{i,t}\), and its target headcount, \(\bar{L}_{i,t}\). The key friction is that effective labor input, \(L_{i,t} = \mu _{i,t} \bar{L}_{i,t}\), where \(0 \leq \mu _{i,t}\leq 1\) is the idiosyncratic ‘job filling’ shock:

\[ \begin{aligned} \max _{M_{i,t}, \bar{L}_{i,t}} P(Y_{it})F_{t}(M_{it},\mu _{i,t}\bar{L}_{it},\Omega _{it})&-x_{t}^{M}M_{it}-w(\bar{L}_{it})\mu _{i,t}\bar{L}_{it}-C(\bar{L}_{it}) \end{aligned} \]

Output (\(Y_{it}\)) is produced using two inputs, intermediate inputs (\(M_{it}\)) and effective labor (\(L_{it}=\mu _{i,t}\bar{L}_{it}\)), according to the common production function \((F_{t})\). \(\Omega _{it}\) denotes idiosyncratic productivity shocks. The firm faces a demand curve that determines the price (\(P_{it}\)) as a function of its output level.⁶ On the cost side, intermediate inputs are the flexible input (á la Gandhi et al. (2020)) and represent materials, supplies, and other variable inputs excluding labor and are purchased at price, \(x_{t}^{M}\).

However, the labor decision is more nuanced than in standard models. Firms must distinguish between two labor-related quantities: the number of positions they create or workers they aim to hire (\(\bar{L}_{it}\)) and the effective labor input that actually contributes to production (\(L_{i,t} = \mu _{i,t}\bar{L}_{it}\)). This distinction captures hiring frictions and new-hire training costs. For labor costs, the wage function \(w(\bar{L}_{it})\) determines the base wage costs and may reflect the firm-specific labor supply curve and its influence on wages, particularly in markets where it has significant hiring power. The function \(C(\bar{L}_{it})\) captures additional labor-related expenses such as recruitment, vacancy posting, training, signing bonuses, etc.

Unlike the typical firm problem with only direct wage costs of labor, the marginal cost of labor in this framework is given by:

\[ \begin{aligned} MCL_{it}=w(\bar{L}_{it})+\underbrace{\frac{w_{\bar{L}}\left (\bar{L}_{it}\right)L_{i,t}}{\mu _{i,t}}}_{\text{Wage pressure}}+\underbrace{\frac{C_{\bar{L}}\left (\bar{L}_{it}\right)}{\mu _{i,t}}}_{\text{Hiring/training costs}} = w(\bar{L}_{it}) + \lambda _{i,t}\left (\bar{L}_{it}, \mu _{i,t}\right) \end{aligned} \]

where \(\lambda _{i,t}^L=\lambda _{i,t}\left (\bar{L}_{it}, \mu _{i,t}\right)\) is the additional marginal cost of hiring workers over and above the wage costs. Changes in \(MCL_{it}\) are a key object for many firm decisions. For example, from the first order condition (FOC) with respect to labor input \(L_{it}\), one can show that under flexible prices the pass-through rate of changes in marginal labor cost to prices is one:

\[ \begin{aligned} \log \left (P_{it}\right) & = \underbrace{\log \left (\frac{\epsilon _{it}}{\epsilon _{it}-1}\right)}_{\text{Markup} (\eta _{it})} + \underbrace{\log \left (w\left (\bar{L}_{it}\right)+\lambda _{it}^{L}\right)}_{\text{Marginal cost of labor}} - \underbrace{\log \left (Y_{it}/L_{it}\right)}_{\text{Labor productivity}} - \underbrace{\log \left (\alpha _{it}^{L}\right)}_{\text{Output elasticity}}, \\ \Delta \log \left (P_{it}\right) & = \Delta \log \left (w\left (\bar{L}_{it}\right)+\lambda _{it}^{L}\right) - \Delta \log \left (Y_{it}/L_{it}\right) - \Delta \log \left (\alpha _{it}^{L}\right) - \Delta \log \left (\eta _{it}\right) \end{aligned} \]

where \(\epsilon _{it}\) is the price elasticity of demand and \(\alpha _{it}^{L}=\frac{\partial Y_{it}}{\partial L_{it}}\frac{L_{it}}{Y_{it}}\) is the output elasticity of labor.

Despite its central role in firm’s decisions, measuring changes in marginal labor costs presents a challenge since financial data does not provide such detailed breakdowns. To this end, we manipulate the optimality conditions with respect to intermediate and labor inputs:

\[ \begin{aligned} \frac{\partial Y_{i,t}/\partial L_{i,t}}{\partial Y_{i,t}/\partial M_{i,t}} &= \frac{MCL_{i,t}}{x^M_t} \Rightarrow \frac{x_{t}^{M}M_{it}}{P_{i,t}Y_{it}} = \frac{\left (w\left (\bar{L}_{it}\right)+\lambda _{it}^{L}\right)L_{it}}{P_{i,t}Y_{it}} \cdot \frac{\alpha _{it}^{M}}{\alpha _{it}^{L}} \Rightarrow \\ \underbrace{\Delta \log s_{it}^{M}}_{\text{$M$-share}} & = \underbrace{\Delta \log \left (w\left (\bar{L}_{it}\right)+\lambda _{it}^{L}\right)}_{\text{Marginal cost of labor}} - \underbrace{\Delta \log \frac{P_{i,t}Y_{it}}{L_{it}}}_{\text{Revenue per worker}} + \underbrace{\Delta \log \frac{\alpha _{it}^{M}}{\alpha _{it}^{L}}}_{\text{Output elasticities}}, \end{aligned} \]

where \(\alpha _{it}^{M}=\frac{\partial Y_{it}}{\partial M_{it}}\frac{M_{it}}{Y_{it}}\) is the output elasticity of intermediate input and \(s_{it}^{M}\) is the revenue share of intermediate input cost.⁷

Our theoretical framework provides a clear, testable prediction: when a firm faces rising marginal labor costs (our unobservable \(\mu _{it}\) shock), it will substitute away from labor and towards other variable intermediate inputs. This substitution will, all else equal, increase the share of revenue spent on intermediate inputs (\(s^M_{it}\)). This relationship is the key to estimating our measure. While Compustat does not report marginal labor costs, our textual analysis of earnings calls provides a rich, high-frequency proxy for the pressures driving them (\(\Lambda _{it}\)).

We can therefore quantify the economic significance of these discussions by directly estimating their impact on the intermediate inputs share. This leads to our core empirical specification, which rearranges the first-order condition from Equation (4):

\[ \Delta \log \left (w\left (\bar{L}_{it}\right)+\lambda _{it}^{L}\right)=f(\Lambda _{it})+\varsigma _{it}, \]

where \(\Lambda _{it}=[\Lambda _{it}^{1},\Lambda _{it}^{2},...,\Lambda _{it}^{8}]^{\prime}\) is an array of firm \(i\)’s exposure to topic \(k\) of labor-related issues in quarter \(t\)—\(\Lambda _{it}^{k}\)—, which is measured as the count of instances of keywords normalized by the length of the call (see Section 2). Therefore, in our empirical specification we project changes in firms’ intermediate input revenue shares onto their exposure to labor topics and estimate \(\omega _{it}=f(\Lambda _{it})\) in the below equation:

\[ \underbrace{\Delta \log s_{it}^{M}}_{\text{M-share}}=\underbrace{f(\Lambda _{it})}_{\omega _{it}}-\underbrace{\Delta \log \frac{P_{i,t}Y_{it}}{L_{it}}}_{\text{Revenue per worker}}+\epsilon _{it}, \]

where \(\epsilon _{i,t}\) denotes the error term that absorbs changes in output elasticities and measurement error as well as possible variation in utilization of inputs.

This regression approach converts our textual data from earnings calls into quantifiable estimates of labor cost pressures. It provides an economically grounded method for condensing the complex, multidimensional information in earnings calls into a single meaningful measure. By using the regression coefficients as weights, we create a composite measure that captures how different labor-related discussions contribute to overall firm costs. This approach yields a comprehensive labor cost pressure measure that weights each topic’s mention according to its estimated impact on firms’ cost structures, combining the rich qualitative information from earnings calls into a single, economically interpretable metric. Finally, the theoretical framework also mitigates overfitting—a pervasive problem in high-dimensional textual analysis (Gentzkow et al., 2019).

One might consider an alternative ‘residual approach’, where labor cost pressures are defined as the unobserved component from a regression of intermediate inputs’ revenue share on revenue per worker. However, such an approach suffers from a critical drawback. While our method uses external, independently measured topic signals on the right hand side and only requires that this signal is not systematically correlated with other unobservables after including fixed effects. The residual approach requires a much stronger and less plausible assumption. Specifically, it assumes that any variation in intermediate inputs’ revenue share not explained by revenue per worker is variation in labor cost pressures, thereby mechanically conflating the signal with all sources of noise and misspecification.

4 Estimating labor cost pressures, \(\omega _{it}\)

4.2 Labor cost pressures (\(\omega _{it}\)) estimates

We regress firms’ revenue shares of variable intermediate inputs on labor-related topics while controlling for firm and time fixed effects. We flexibly control for employment growth and sales growth to account for measurement error in reporting employment. To allow for flexible output elasticities—which could vary by industry and firm and over time (Hubmer et al., 2024)—we control for industry-year and firm fixed effects.⁹ In particular, we estimate equation 6 at the annual level using the following regression specification.

\[ \Delta \log s_{i,t}^{M}=\underbrace{\sum _{k}\beta _{k}^{topic}\Lambda _{i,t}^{k}}_{\omega _{it}}+\beta _{1}\Delta log(Emp_{i,t})+\beta _{2}\Delta log(Sales_{i,t})+\delta _{jt}+\varepsilon _{it}. \]

\(\Delta log(Intermediate \ Input/Sales)_{i,t}\)
	(1)	(2)	(3)	(4)	(5)
Labor Costs Higher\(_{i,t}\)	4.270***	3.981***	3.427***	3.580***	3.854***
	(0.613)	(0.604)	(0.675)	(0.686)	(0.731)
Labor Costs Lower\(_{i,t}\)	–6.075***	–6.437***	–6.739***	–6.685***	–6.497***
	(0.973)	(0.971)	(1.008)	(1.044)	(1.014)
Headcount Higher\(_{i,t}\)	–0.788	–0.529	0.263	0.392	0.882
	(0.724)	(0.727)	(0.733)	(0.730)	(0.930)
Headcount Lower\(_{i,t}\)	–0.981	–0.365	–0.188	0.444	–0.764
	(1.032)	(1.029)	(1.038)	(1.075)	(1.162)
Labor Shortage\(_{i,t}\)	2.042	2.446	1.234	1.231	2.019
	(1.808)	(1.890)	(2.118)	(2.154)	(2.486)
Labor Efficiency Higher\(_{i,t}\)	–1.061	–1.838	–2.899	–2.883	–3.636
	(2.821)	(2.774)	(2.928)	(3.011)	(3.146)
Labor Efficiency Lower\(_{i,t}\)	6.885**	7.447**	7.494**	8.318**	1.373
	(3.291)	(3.287)	(3.464)	(3.608)	(3.398)
Labor Agreement\(_{i,t}\)	1.013	0.813	–0.425	–0.696	–0.441
	(1.102)	(1.082)	(1.112)	(1.167)	(1.247)
Residual category\(_{i,t}\)	0.374	0.286	0.520*	0.556*	1.056**
	(0.282)	(0.285)	(0.312)	(0.333)	(0.421)
\(R^2\)	0.050	0.060	0.099	0.099	0.223
N	23,790	23,790	23,714	21,500	21,240
Baseline Controls	Y	Y	Y	Y	Y
Time FE	N	Y	Y	Y	Y
Industry-Year FE	N	N	Y	Y	N
Sentiment and Risk	N	N	N	Y	Y
Firm FE	N	N	N	N	Y

Table 2: Estimation of labor cost pressures: Labor topics and variable input cost share, at firm x year level

Notes: The table shows regression of changes in on labor topics. Each observation denotes a firm i and year t. To construct labor topic observations at the firm x year level we take averages across all quarters in a year. All specifications include controls for yearly changes in sales and employment. Columns (4) and (5) include controls for risk and sentiment. We exclude firms in financial and administrative industries (NAICS 52, 53 and 56). We only include firm level data till 2019 to exclude Covid-19 pandemic period. Standard errors are clustered by firm.

where \(s_{i,t}^{M}\) is the share of intermediate input and \(Emp_{i,t}\) and \(Sales_{i,t}\) are employee count and sales reported by firm \(i\) in year \(t\). Each topic score for topic \(k\) by firm \(i\) in year \(t\) is constructed by averaging over the topic scores calculated quarterly for each earnings conference call. \(\delta _{j,t}\) denotes various controls such as firm-level risk and sentiment scores, and industry, firm, and time fixed effects.

In this section we show the regression results when we use Demirer (2022)’s material measure (i.e., the sum of COGS and SGA minus DP and wage expenditures). The regression results for other measures of intermediate inputs—COGS minus imputed wage expenditures or COGS and material inventory additions—are remarkably similar (Table A.5). Table 2 shows the regression coefficients that relate labor topics to firms’ cost structures. We present five different specifications with progressively more stringent controls, ranging from a basic specification to one that includes firm fixed effects, time fixed effects, and additional controls for risk and sentiment. Using these coefficients, we calculate the estimated change in marginal costs of labor using:

\[ \omega _{it}=\sum _{k}\hat{\beta}_{k}^{topic}\Lambda _{k,i,t} \]

Magnitudes and statistical significance of these coefficients exhibit interesting variations. Specifically, discussions of higher labor costs are consistently strongly associated with increases in the intermediate input-to-sales ratio. The estimated coefficients range from 4.270 (s.e. = 0.613) in the baseline model without any controls to 3.854 (s.e.= 0.731) in our most restrictive specification that includes firm and time fixed effects. In our most restrictive specification, the coefficient remains highly statistically significant at the 1% level. Similarly, mentions of lower labor costs are linked to significant negative coefficients, ranging from -6.075 (s.e.=0.973) to -6.497 (s.e.=1.014).

Appendix table A.6 shows regressions with individual topics. The table shows that individually discussion of higher and lower labor costs, and decreasing headcounts are significantly associated with changes in intermediate input-to-sales ratio. However, these associations are sufficiently summarized in the discussion of higher and lower labor cost pressures and therefore are not statistically significant in our baseline table.

The stability of key coefficients across specifications for labor costs discussions, suggests these relationships are robust and not driven by unobservables at the firm, industry, or time level.

Figure 4: Labor Cost Pressure Index (\(\bar{\omega}_{t}\))

Notes: The figure shows the average change in estimated marginal cost of labor across firms by quarter

along with bootstrapped confidence intervals. The estimation uses coefficients from Table

2

and topic specific

scores by firm and by quarter. Bootstrapped confidence intervals are calculated by randomly leaving out

10% of the sample one at a time for the estimation procedure for 200 samples.

5 Labor Cost Pressures and Inflation Pass-through

5.1 Estimating pass-through to PPI inflation

We begin by quantifying the pass-through of labor cost pressures (\(\omega _{it}\)) to Producer Price Index (PPI) inflation. To do so, we specify an equation by combining the marginal cost calculated at the firm level, and then taking a sales weighted average at the industry level. From Equation 5 with a log-linearized version of the pricing equation 3:

\[ \hat{P}_{i,t}=\beta _{p}\omega _{it}-\widehat{\left (Y_{it}/L_{it}\right)}-\hat{\alpha}_{it}^{L}-\hat{\eta}_{it}, \]

\[ \begin{aligned} \hat{P}_{n,t} & =\beta _{p}\sum _{i\in n}\theta _{i,t}\omega _{it}-\widehat{\left (Y_{nt}/L_{nt}\right)}-\hat{\alpha}_{nt}^{L}-\hat{\eta}_{nt}, \end{aligned} \]

where \(log(P_{n,t})=\sum _{i,t}\theta _{i,t}log(P_{i,t})\) is the aggregate price index for industry \(n\), and \(\theta _{i,t}\) denotes the sales share of firm \(i\) in industry \(n\) at time \(t\).

To estimate \(\beta _{p}\) and understand the dynamic response of prices to these pressures, we employ the following Jordà local projection specification:

\[ log(PPI_{n,t+h})-log(PPI_{n,t})=\alpha _{h}+\beta _{h}\bar{\omega}_{nt}+\sum _{k=1}^{3}\gamma _{k}\Delta log(PPI_{n,t-k})+\delta _{n,h}+\delta _{t,h}+\epsilon _{n,t+h} \]

Figure 7: Dynamic response of industry inflation to labor cost pressures

Notes: The figure plots estimated pass-through of labor cost pressures to PPI inflation (

\(\beta _h\)

) for horizons

\(h=0\)

to

\(8\)

. Shaded areas represent

95% confidence intervals. Standard errors are clustered by industry. The regression is weighted by the number of firms observed

within the industry. We exclude financial and administrative industries (NAICS 52, 53 and 56). Inflation is winsorized at the

2nd and 98th percentiles.

Figure 8: Pass-through of labor cost pressures by Industry

Notes: The figure plots estimated pass-through coefficients at horizon

\(h=4\)

of labor cost pressures to PPI inflation by 2-digit NAICS

industry. The plot shows estimates and 95% confidence intervals. Standard errors are clustered by industry.

In our analysis we use a panel of 2,832 industry-quarter observations (Table A.7). Figure 7 shows that these labor cost pressures take about 6-8 quarters to pass-through 86 percent of the increase in labor cost pressures to PPI inflation. Table A.7 shows regression coefficients with various specifications (with and without fixed effects), and shows a consistently positive and significant relationship between industry-level labor cost pressures and subsequent PPI inflation within the next four quarters. The coefficient on the labor cost pressure measure ranges from 0.387 to 0.493, indicating that a one percentage point increase in labor cost pressures is associated with roughly 0.493 percentage points higher PPI inflation four quarters later.

We check that these results are robust to calculating intermediate input shares using alternative methods, and that they are robust to calculating topics using the unsupervised approach. Table A.9 shows that these patterns remain largely unchanged with other measures of labor cost pressures calculated using alternative proxies for intermediate inputs or calculated using unsupervised approach. Figure A.4 shows that the pass-through estimates are quite stable for any chose number of principal components greater than 6.

We compare the pass-through of labor cost pressures against the pass-through of changes in ECI at the industry level in figure A.3 with the same Jorda projection specification. We find no significant pass-through of ECI on to PPI inflation.

However, this average effect masks significant heterogeneity. Figure 8 plots the interaction of the labor cost pressure coefficient with industry dummies. Industries with high labor intensity—such as Information, Wholesale Trade, and Accommodation and Food Services—exhibit significantly higher pass-through. Conversely, the pass-through for the manufacturing sector is near zero. These results align with Heise et al. (2022), who find significant wage-price pass-through in services but not in manufacturing. While they attribute the manufacturing disconnect to import competition and market concentration, our findings in Section 6 suggest that investment in automation provides an alternative explanation: manufacturing firms mitigate cost pressures through capital-labor substitution rather than price increases.

5.2 Aggregate Labor Cost Pressures and PCE Inflation

Having established the link between our measure and producer prices, we now examine its implications for aggregate consumer inflation. We construct an aggregate Labor Cost Pressure Index (\(\bar{\omega}_{t}\)) by weighting our industry-level measures using Personal Consumption Expenditures (PCE) industry weights.

Figure 9 plots this aggregate index alongside core PCE inflation. The correlation is striking. Our index captures the short-lived disinflation during the Great Recession and the rapid surge during the post-pandemic recovery. Quantitatively, assuming full pass-through to inflation, our measure can explain 19.5 percentage points of the 29.3 percentage point cumulative surge in inflation during the post-pandemic period (2021-22).

Figure 9: Labor Cost Pressure Index (\(\bar{\omega}_{t}\)) and PCE Inflation

Notes: The figure shows our labor cost pressure index and the PCE inflation. To aggregate with PCE weights, we use industry

level aggregate at the NAICS 3 digit level and then aggregate up to quarterly level using PCE weights for each industry.

6 How do firms respond to labor cost pressures?

Our finding that manufacturing exhibits near-zero inflation pass-through from labor cost pressures suggests firms adjust through margins other than prices. We now exploit the firm-level nature of our measure to examine whether labor cost pressures trigger automation and productivity gains—particularly in industries where routine tasks make capital-labor substitution feasible.

Capital–labor substitution in production is central for many questions in economics, such as factor income shares (e.g., Karabarbounis and Neiman (2014)) or earnings distribution (e.g., Krusell et al. (2000)). Most models of labor demand imply that when firms face higher labor costs they adopt automation technologies (e.g., Acemoglu and Autor (2011); Acemoglu and Restrepo (2022a); Leduc and Liu (2024)).

We test this prediction using firm-quarter observations from Compustat, analyzing how labor cost pressures affect investment, R&D spending, and productivity. We use the following specification:

\[ log(CapEx_{i,t,t+4})=\alpha +\beta \omega _{i,t}+\gamma log(assets_{i,t-1})+\chi _{i,t}+\delta _{i}+\delta _{t}+v_{t}, \]

where \(\omega _{it}\) denotes labor cost pressures at the firm level observed in quarter t. \(log(CapEx_{i,t,t+4})\) denotes log capital expenditures between quarter t and t+4. \(\delta _{i}\)and \(\delta _{t}\) are firm and time fixed effects. We control for risk and sentiment in expressed in earnings conference calls following Hassan et al. (2024a). We focus on heterogeneity across industries with different levels of routine manual task intensity (Table 7 Panel A). The baseline specification (column 1) indicates that a percentage point increase in labor cost pressures is associated with a 2.58% increase in capital expenditure. The inclusion of various fixed effects and controls helps isolate the effect of labor cost pressures from other factors that might influence investment decisions, such as industry trends, macroeconomic conditions, and firm-specific characteristics.

Columns 2-3 reveal interesting heterogeneity across industry types based on their routine manual task intensity. High routine manual (HRM) industries show the strongest investment response to labor cost pressures (4.20%), followed by low routine manual (LRM) industries (2.24%), while medium routine manual (MRM) industries show the smallest response (1.89%). This pattern suggests that firms in industries with high routine manual task content are more likely to respond to labor cost pressures by increasing capital investment, consistent with greater opportunities for automation and capital-labor substitution in these industries.

Table 7:

Labor Cost Pressures and Investment, Firm x quarter Level

Notes: The table shows regression of log(Capital expenditures) reported by firm i between quarters t+1 and t+4 on labor cost pressures calculated for firm \(i\) in quarter \(t\). Each observation denotes a firm \(i\) and quarter \(t\). All columns include controls for risk, sentiment and log assets. We exclude firms in financial and administrative industries (NAICS 52, 53 and 56). Standard errors are clustered by firm.

We next examine how firms adjust their R&D expenditures in response to labor cost pressures, using a similar framework as the capital expenditure analysis (Table 7 Panel B). For the full sample of firms (columns 1-3), labor cost pressures show a positive and significant relationship with R&D spending. The effect is substantial, with a one percentage point increase in labor cost pressures associated with a 0.56% increase in R&D expenditure in the baseline specification. We find similar results to those observed for capital expenditures: firms in high routine manual industries invest in R&D at 7 times the rate compared to low routine manual industries. This finding, combined with the heterogeneous responses across industry types, indicates that firms’ technological capabilities and industry characteristics significantly influence their choice between capital investment and R&D as responses to labor cost pressures.

Do these investments translate into productivity gains? Figure 11 presents an event study analyses showing how labor productivity (measured as revenue per employee) responds to labor cost pressure episodes from 4 years before to 5 years after the shock, with confidence intervals shown by the dotted lines. Panel A, focusing on HRM industries, reveals a clear pattern of productivity gains following labor cost pressure episodes. While there is no significant pre-trend before the shock (years -4 to 0), productivity begins to increase notably around year 1 and continues to rise, reaching a peak of nearly 1% higher productivity by year 3. The effect remains positive and statistically significant through year 5, suggesting persistent productivity improvements in these industries following labor cost pressures.

In contrast, Panel B shows that low routine manual industries exhibit no significant productivity response to labor cost pressure episodes. The point estimates fluctuate around zero after increases in labor cost pressures, and the confidence intervals consistently include zero. This stark difference between HRM and LRM industries aligns with the earlier findings on investment and automation responses, suggesting that firms in HRM industries are more successful at converting their technological responses to labor cost pressures into actual productivity gains, likely through successful automation and capital-labor substitution.

These findings explain our earlier result that manufacturing shows near-zero inflation pass-through. When labor cost pressures hit, manufacturing firms—concentrated in HRM industries—respond by automating rather than raising prices. The capital deepening and productivity gains offset the cost shock. In contrast, service sectors with fewer automation opportunities must pass costs through to prices, generating the heterogeneous pass-through documented in Figure 8. The firm-level evidence thus provides a mechanism for the industry-level inflation patterns, highlighting how technological substitution possibilities shape the inflation consequences of labor market tightness.

7 Conclusion

Labor cost pressures have long been recognized as a fundamental driver of inflation since Phillips (1958). Yet, measuring marginal labor costs remains challenging because financial statements lack such detailed breakdowns. In this paper, we developed a novel methodology to address this challenge by combining textual analysis of earnings calls with economic theory. Our theory-based measurement approach aggregates multidimensional rich qualitative information into a single quantitative time-varying firm-level measure. The theoretical structure also prevents overfitting—a common pitfall when working with high-dimensional textual data.

We use this measure to examine the pass-through from labor cost pressures to inflation. First, when aggregated using PCE weights, our index outperforms conventional slack variables in forecasting core PCE inflation. Second, we find significant but heterogeneous pass-through to industry-level PPI inflation: the effect is strongest in services and near-zero in manufacturing. Third, exploiting firm-level granularity, we show that labor cost pressures prompt increased investment in industries with high share of routine manual workers, generating productivity gains that help explain manufacturing’s muted price response.

These results advance our understanding of inflation dynamics and offer practical tools for real-time monitoring. Policymakers could use our high-frequency measure to anticipate inflationary pressures earlier than traditional indicators allow. Our firm-level evidence reveals that labor cost pressures accelerate automation, particularly where capital-labor substitution is technologically feasible, with implications for both productivity growth and the distributional effects of tight labor markets. More broadly, our methodology demonstrates how economic theory can discipline textual analysis to measure hard-to-quantify firm-level variables, a framework applicable beyond labor costs.

8 References

Acemoglu, D. and Autor, D. H. (2011). Skills, tasks and technologies: Implications for employment and earnings. Handbook of Labor Economics, 4, 1043–1171.

— and Restrepo, P. (2022a). Demographics and automation. The Review of Economic Studies, 89 (1), 1–44.

— and — (2022b). Tasks, automation, and the rise in us wage inequality. Econometrica, 90 (5), 1973–2016.

Ahn, H. J., Cook, T. R. and Doh, T. (2025a). Bargaining over words? text analysis of a model of monetary policy by a committee. Federal Reserve Bank of Kansas City Working Paper.

—, —, —, Kastritis, E. and Wedewer, J. (2025b). Text sentiment about monetary policy. Federal Reserve Bank of Kansas City Working Paper No. RWP25-18.

Araujo, D., Bokan, N., Comazzi, F. and Lenza, M. (2025). Word2prices: embedding central bank communications for inflation prediction.

Arold, B. W., Ash, E., MacLeod, W. B. and Naidu, S. (2024). Do words matter? the value of collective bargaining agreements. Center for Law & Economics Working Paper Series, 6.

Ashwin, J., Kalamara, E. and Saiz, L. (2024). Nowcasting euro area gdp with news sentiment: a tale of two crises. Journal of Applied Econometrics, 39 (5), 887–905.

Atkeson, A., Ohanian, L. E. et al. (2001). Are phillips curves useful for forecasting inflation? Federal Reserve bank of Minneapolis quarterly review, 25 (1), 2–11.

Bagga, S., Mann, L. F., Şahin, A. and Violante, G. L. (2025). Job amenity shocks and labor reallocation. Tech. rep., National Bureau of Economic Research.

Baker, S. R., Bloom, N. and Davis, S. J. (2016). Measuring economic policy uncertainty. The quarterly journal of economics, 131 (4), 1593–1636.

Barnichon, R. and Shapiro, A. H. (2024). Phillips meets beveridge. Journal of Monetary Economics, 148, 103660.

Berger, D., Herkenhoff, K., Kostèl, A. R. and Mongey, S. (2024). An anatomy of monopsony: Search frictions, amenities, and bargaining in concentrated markets. NBER Macroeconomics Annual, 38 (1), 1–47.

—, — and Mongey, S. (2022). Labor market power. American Economic Review, 112 (4), 1147–1193.

Bilal, A., Engbom, N., Mongey, S. and Violante, G. L. (2022). Firm and worker dynamics in a frictional labor market. Econometrica, 90 (4), 1425–1462.

Blanco, A., Boar, C., Jones, C. and Midrigan, V. (2024). Non-linear inflation dynamics in menu cost economies. w32094.

Borusyak, K. and Jaravel, X. (2021). The distributional effects of trade: Theory and evidence from the united states. Tech. rep., National Bureau of Economic Research.

Buehlmaier, M. M. and Whited, T. M. (2018). Are financial constraints priced? evidence from textual analysis. The Review of Financial Studies, 31 (7), 2693–2728.

Chen, H., Cheng, Y., Liu, Y. and Tang, K. (2026). Teaching economics to the machines. NBER Working Paper.

De Loecker, J., Eeckhout, J. and Unger, G. (2020). The rise of market power and the macroeconomic implications. The Quarterly journal of economics, 135 (2), 561–644.

— and Warzynski, F. (2012). Markups and firm-level export status. American economic review, 102 (6), 2437–2471.

De Ridder, M., Grassi, B., Morzenti, G. et al. (2022). The hitchhiker’s guide to markup estimation.

Demirer, M. (2022). Production function estimation with factor-augmenting technology: An application to markups.

Domash, A. and Summers, L. H. (2022). How tight are US labor markets? Tech. rep., National Bureau of Economic Research.

Drechsel, T., McLeay, M. and Tenreyro, S. (2019). Monetary policy for commodity booms and busts.

Fitzgerald, T. J., Nicolini, J. P. et al. (2014). Is there a stable relationship between unemployment and future inflation?: Evidence from US cities. Citeseer.

Gabaix, X. (2011). The granular origins of aggregate fluctuations. Econometrica, 79 (3), 733–772.

Gagliardone, L., Gertler, M., Lenzu, S. and Tielens, J. (2023). Anatomy of the Phillips curve: micro evidence and macro implications. Tech. rep., National Bureau of Economic Research.

Galí, J. (2015). Monetary policy, inflation, and the business cycle: an introduction to the new Keynesian framework and its applications. Princeton University Press.

Galı, J. and Gertler, M. (1999). Inflation dynamics: A structural econometric analysis. Journal of monetary Economics, 44 (2), 195–222.

Gandhi, A., Navarro, S. and Rivers, D. A. (2020). On the identification of gross output production functions. Journal of Political Economy, 128 (8), 2973–3016.

Gentzkow, M., Kelly, B. and Taddy, M. (2019). Text as data. Journal of Economic Literature, 57 (3), 535–74.

Gormsen, N. J. and Huber, K. (2024). Firms’ Perceived Cost of Capital. Tech. rep., National Bureau of Economic Research.

Gorodnichenko, Y., Pham, T. and Talavera, O. (2021). Social media, sentiment and public opinions: Evidence from# brexit and# uselection. European Economic Review, 136, 103772.

—, — and — (2023). The voice of monetary policy. American Economic Review, 113 (2), 548–584.

Gosselin, M.-A. and Taskin, T. (2024). Textual Economic Indicators for Canada and the United States: Insights From Earnings Calls. Tech. rep., Bank of Canada Staff Discussion Paper.

Graetz, G. and Michaels, G. (2018). Robots at work. Review of economics and statistics, 100 (5), 753–768.

Hall, R. E. (1988). The relation between price and marginal cost in us industry. Journal of political Economy, 96 (5), 921–947.

— and Mueller, A. I. (2018). Wage dispersion and search behavior: The importance of nonwage job values. Journal of Political Economy, 126 (4), 1594–1637.

Harford, J., He, Q. and Qiu, B. (2024). Firm-level labor-shortage exposure. Available at SSRN 4410126.

Hassan, T. A., Hollander, S., Kalyani, A., van Lent, L., Schwedeler, M. and Tahoun, A. (2024a). Economic Surveillance using Corporate Text. Tech. rep., National Bureau of Economic Research.

—, —, Van Lent, L. and Tahoun, A. (2019). Firm-level political risk: Measurement and effects. The Quarterly Journal of Economics, 134 (4), 2135–2202.

—, —, van Lent, L. and Tahoun, A. (2023). Firm-level exposure to epidemic diseases: Covid-19, SARS, and H1N1. Review of Financial Studies, 34 (6), 2409–2449.

—, —, — and — (2024b). The global impact of Brexit uncertainty. Journal of Finance, 79, 413–458.

Hazell, J., Herreno, J., Nakamura, E. and Steinsson, J. (2022). The slope of the phillips curve: evidence from us states. The Quarterly Journal of Economics, 137 (3), 1299–1344.

Heise, S., Karahan, F. and Şahin, A. (2022). The missing inflation puzzle: The role of the wage-price pass-through. Journal of Money, Credit and Banking, 54 (S1), 7–51.

Howes, C., i Carreras, M. D., Coibion, O. and Gorodnichenko, Y. (2026). How Monetary Policy Is Made: Lessons from Historical FOMC Discussions. Tech. rep., National Bureau of Economic Research.

Hubmer, J., Chan, M., Ozkan, S., Salgado, S. and Hong, G. (2024). Scalable versus Productive Technologies. Tech. rep., Penn Institute for Economic Research, Department of Economics, University of Pennsylvania.

Hwang, H.-s., Mortensen, D. T. and Reed, W. R. (1998). Hedonic wages and labor market search. Journal of Labor Economics, 16 (4), 815–847.

Karabarbounis, L. and Neiman, B. (2014). The global decline of the labor share. The Quarterly journal of economics, 129 (1), 61–103.

Krusell, P., Ohanian, L. E., Ríos-Rull, J.-V. and Violante, G. L. (2000). Capital-skill complementarity and inequality: A macroeconomic analysis. Econometrica, 68 (5), 1029–1053.

Leduc, S. and Liu, Z. (2024). Automation, bargaining power, and labor market fluctuations. American Economic Journal: Macroeconomics, 16 (4), 311–349.

Moscarini, G. and Postel-Vinay, F. (2012). The contribution of large and small employers to job creation in times of high and low unemployment. American Economic Review, 102 (6), 2509–2539.

Ni, J., Abrego, G. H., Constant, N., Ma, J., Hall, K. B., Cer, D. and Yang, Y. (2021). Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. arXiv preprint arXiv:2108.08877.

Pennington, J., Socher, R. and Manning, C. D. (2014). Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532–1543.

Phillips, A. W. (1958). The relation between unemployment and the rate of change of money wage rates in the united kingdom, 1861-1957. economica, 25 (100), 283–299.

Powell, J. H. (2018). Monetary policy and risk management at a time of low inflation and low unemployment. Business Economics, 53, 173–183.

Reimers, N. and Gurevych, I. (2019). Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084.

Rosen, S. (1986). The theory of equalizing differences. Handbook of labor economics, 1, 641–692.

Song, W. and Stern, S. (2024). Firm inattention and the efficacy of monetary policy: A text-based approach. Review of Economic Studies, p. rdae102.

Stock, J. H. and Watson, M. W. (1999). Forecasting inflation. Journal of monetary economics, 44 (2), 293–335.

— and — (2008). Phillips curve inflation forecasts.

9 Appendix Tables and Figures

Table A.1:

Top 20 Firms by Labor Discussion Intensity

Notes: This table presents the 20 firms with the highest average labor discussion intensity across all quarterly earnings calls in the sample. Labor count represents the percentage of sentences in earnings call transcripts that discuss labor-related topics, averaged across all available quarters for each firm. The sample includes U.S. public firms with at least 20 quarterly observations. Industry classifications are based on 3-digit NAICS codes.

Table A.2:

Top 20 keywords by topic

Notes: This table shows top 20 keywords used for each of our labor topics.

Table A.3:

Discussion of labor topics and wage growth of new hires, at industry level

Notes: The table shows regression of changes in on discussion of labor topics by industry n in quarter t. Earnings denotes earnings for new hires observed in the quarterly workforce indicates aggregated over industry n at time t. Industry is at the NAICS 3-digit level. Each observation denotes a industry n and year t. To construct labor topic observations at the industry x year level we take averages across all quarters in a year. We exclude firms in financial and administrative industries (NAICS 52, 53 and 56). Topic variables are standardized. Regression is weighted by number of earnings calls in the industry. Standard errors are clustered by industry.

Table A.4:

Summary Statistics

Notes: This table reports summary statistics for key variables at four levels of aggregation. Panel A shows firm-year level data for the baseline estimation sample (2002–2019), which includes U.S.-headquartered firms with at least 5 years of non-missing labor topic data, excluding mining (NAICS 21), finance (NAICS 52), real estate (NAICS 53), and administrative services (NAICS 56). Panel B shows industry-quarter level data at the NAICS 3-digit level, excluding import-dependent industries (NAICS 315, 334, 339) and FIRE industries (NAICS 520–699). Panel C shows aggregate quarterly time series (2002Q1–2024Q4). Labor topic variables measure the percentage of sentences in earnings calls mentioning each topic. Labor Cost Pressures is the fitted value from regressing changes in the intermediate inputs-to-sales ratio on labor topics (Equation 6). All continuous variables are winsorized at the 1st and 99th percentiles except where noted.

Table A.5:

Estimation of labor cost pressures: Labor topics and variable intermediate input cost share calculated using material inventories, at firm x year level

Table A.6:

Firm level regressions, topics entered individually

Notes: Each column reports a separate regression of the dependent variable on a single labor topic, controlling for labor exposure, sentiment, risk, and firm growth controls. All specifications include firm and year fixed effects, and standard errors are clustered by firm. The sample is restricted to years up to 2019.

Figure A.2: Labor Cost Pressure Index (\(\bar{\omega}_{t}\)) across Industries

Figure A.3: Labor Cost Pressure Index (\(\bar{\omega}_{t}\)), Employment Cost Index, and PPI Inflation

Table A.7:

Labor Cost Pressures and Inflation, Industry x time Level

Notes: The table shows regression of PPI inflation observed in industry n between quarters and on estimated changes in MCL for industry in quarter. Industry is one of NAICS 3-digit industries. Estimated changes in MCL for industry in quarter is calculated by taking a sales weighted average of estimated changes in MCL across all firms in Compustat in the industry. We exclude financial and administrative industries (NAICS 52, 53 and 56). Regression is weighted by number of firms observed within the industry. Standard errors are clustered by industry. Inflation is winsorized at 2nd and 98th percentile.

Table A.8:

Variance Decomposition of Labor Cost Pressures

Notes: The table reports results from regressing firm-level labor cost pressures (\(\omega _{it}\)) on various fixed effects and firm-specific controls. Columns (1)–(5) include time, industry \(\times\) time, and firm fixed effects. Columns (6) and (7) include granular fixed effects, where Size and Growth quintiles are constructed based on the pooled distribution of firm employment and lagged stock returns, respectively. Risk and Net Sentiment are text-based measures constructed following Hassan et al. (2024a). Standard errors are clustered at the firm level.

Figure A.4:

Robustness: Phillips Curve Estimation using Unsupervised Methodology (PCA)

Notes: This figure plots the coefficient estimates and 95% confidence intervals from a series of regressions. The dependent variable is PPI inflation observed in industry \(n\) between quarters \(t\) and \(t+4\). The x-axis represents different specifications: Index 0 is the baseline Labor Cost Pressure measure (\(MCL_{n,t}\)), while indices 1 through 50 represent specifications using the first 50 Principal Components derived from the underlying text data. The underlying \(MCL\) is calculated as a PCE-weighted average of estimated changes in labor cost pressures across US-headquartered firms in Compustat that hold earnings conference calls in quarter \(t\). We exclude financial and administrative industries (NAICS 52, 53, and 56). All regressions include industry (NAICS3) and time (Date) fixed effects. Standard errors are robust.

Table A.9:

Industry level Phillips Curve robustness to alternative labor cost pressure measures

Notes The table reports industry level regressions of four quarter ahead year on year PPI inflation on alternative measures of labor cost pressures, using NAICS three digit industries as the cross section. All specifications include industry and quarter fixed effects and are weighted by mean number of firms holding earnings conference calls. Standard errors are robust.

Table A.10:

Robustness of Phillips Curve to alternative labor cost pressure measures

Notes: The table reports OLS estimates of the Phillips Curve specification using alternative measures of labor cost pressures. The dependent variable is PCE inflation over the next four quarters. Column (1) uses the baseline SGA based measure, column (2) the inventory based measure, column (3) the alternative intermediate inputs measure, and column (4) the unsupervised PCA based measure. All regressions control for lagged Michigan inflation expectations. Standard errors are robust.

Figure A.5: Marginal cost of labor estimates calculated using unsupervised methodology

Decomposition using Labor Market Tightness

Decomposition using Unemployment Rate

Figure: Figure A.6: Decomposition of Labor Cost Pressures: Systematic vs. Idiosyncratic

Table A.11:

Inflation Decomposition: Systematic vs. Idiosyncratic Pressures, without lagged inflation control

Notes: The table reports results from regressing PCE inflation on the systematic and idiosyncratic components of the Labor Cost Pressure index. The Systematic Component is constructed by aggregating firm-level predicted values based on aggregate slack (Unemployment or Tightness). The Idiosyncratic Component is constructed by aggregating firm-level residuals. All independent variables are standardized. Standard errors are robust.

10 Data Appendix