## Mining the Gap: Extracting Firms’ Inflation Expectations From Earnings Calls Working Paper No. WP/2023/202

## Source details

**Canonical URL:** [Mining the Gap: Extracting Firms’ Inflation Expectations From Earnings Calls Working Paper No. WP/2023/202](https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023202-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2023/english/wpiea2023202-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2023/english/wpiea2023202-print-pdf.pdf.json)

---

### Motivation and overview
- Inflation reached decades-high levels in 2022 and persistent core inflation; households’ and firms’ inflation expectations can be crucial since they are primary price and wage setters.
- Earnings calls provide a high-frequency, firm-level textual source to proxy firms’ inflation expectations and attention to the central bank, enabling replication and alignment with balance sheet data via Compustat.
- Coverage claims:
  - Sample: more than one thousand firms since 2002 and more than two thousand firms since 2004.
  - ECFIE index encompasses 96.3% of the US market capitalization.
  - Training data: S&P Capital IQ 13,407 US companies earnings call transcripts released in 2022.
  - Index computed using more than 200,000 transcripts of publicly listed companies over 39 countries (26 advanced economies and 13 emerging markets).

### Novel data and methodology (ECFIE and ECFACB)
- ECFIE construction (text-mining + human+GPT guidance):
  - Dictionary selection: top 100 transcripts by frequency of "inflation" and "expectation"; randomly sampled 4,000 sentences; human and GPT classified sentences and GPT provided key words to form the dictionary.
  - Index formula (rescaled for presentation):
    - ECFIE Index_it = 1000 × (P Sentences with Inflation Expectations_it) / (P Sentences_it)
  - Human+GPT-guided dictionary improves correlations with external measures relative to a naive search.
- ECFACB (attention to the central bank) construction:
  - Attention to the FED Index_it = 1000 × (P Sentences with Central Bank words_it) / (P Sentences_it)
  - Attention dictionary includes terms like Fed, Fed Funds, monetary policy, central bank, FOMC, quantitative easing, quantitative tightening, monetary easing, monetary tightening, Federal Reserve.
- Methodological claims:
  - Sentence-level, quarterly frequency allows firm-level tracking.
  - Approach combines Generative Pre-trained Transformers (GPT) and human judgment to select dictionary words; NL Analytics used for computation.

### Validation and predictive power (key statistics and correlations)
- Correlations with external measures:
  - Correlation of ECFIE index with Livingston Survey: 0.84
  - Correlation of ECFIE index with CPI-U inflation: 0.8
  - ECFIE vs SoFIE: ρ = 0.9764
  - Disagreement correlation ECFIE vs SoFIE: ρ = 0.7996
  - Appendix Table 1 method vs naive correlations (examples):
    - Livingston Survey: 0.8415 vs 0.7545
    - CPI: 0.8096 vs 0.6707
- Time-horizon cross-correlations with Livingston Survey (Table 3):
  - Next 1M (end of the period): 0.8296
  - Next 6M (end of the period): 0.8339
  - Next 12M (end of the period): 0.8250
  - Following 1Y (average): 0.8415
  - Next 10Y (average): 0.1383
- Predictive power for future inflation:
  - Local projection: π_{t+h} = α_h + β_H π^e_t + ξ_H X_t + ε_{t+h}
  - Finding: "An one-unit increase in ECFIE is associated with an increase in two percent inflation in the US in the first four quarters."
- ECFACB external validation:
  - Correlation between ECFACB and Google Trends measure: 0.557

### Sectoral and firm-level validation
- Sectoral validation:
  - Almost perfect relationship between sectors exposed to inflation surprises (via sectoral returns) and sectors affected by ECFIE.
  - Sector timing differences: Consumer Non-Cyclical, Consumer Cyclical, Basic Materials, Industrials and Energy peaked in 2022Q2; Utilities, Financials, Real Estate, Healthcare and Technology peaked in 2022Q3 or 2022Q4.
- Firm-level statistics (Appendix B.1 — exact values):
  - All Firms (Obs 178479)
    - Inflation Expectation Index: Mean 0.882, SD 2.98, Median 0.00
    - Total Assets (US$million): Mean 1784791865.52, SD 6822.54, Median 3444.98
    - Leverage: Obs 173691, Mean 0.27, SD 0.34, Median 0.23
  - High Inflation Expectations (Obs 89220)
    - Inflation Expectation Index: Mean 1.664, SD 4.01, Median 0.00
    - Total Assets (US$million): Mean 892202176.54, SD 6183.01, Median 544.33
    - Leverage: Obs 87399, Mean 0.29, SD 0.24, Median 0.27
  - Low Inflation Expectations (Obs 89259)
    - Inflation Expectation Index: Mean 0.100, SD 0.65, Median 0.00
    - Total Assets (US$million): Mean 892591554.63, SD 7393.78, Median 207.96
    - Leverage: Obs 86292, Mean 0.24, SD 0.41, Median 0.16
- Stylized facts:
  - Larger firms discuss inflation expectations more than smaller firms; series are highly correlated across size percentiles.
  - Higher-leverage firms discuss inflation expectations more; series are highly correlated across leverage groups.

### Deviations from Full-Information Rational Expectations (FIRE)
- Aggregate and micro evidence show violations of FIRE: firm characteristics and sentiment influence how firms adjust expectations after monetary policy shocks.
- Financial constraint (long-term debt maturing within one year; standardized at sectoral level):
  - Empirical findings:
    - Firms with higher financial constraints increase their inflation expectations by approximately 0.15 units more in response to a monetary policy shock.
    - Average impact of a monetary policy shock: decrease of 0.4 units in the expectation index.
    - For a firm one standard deviation more financially constrained than the sector average just before the shock, the total effect of the shock is weakened by 37%.
  - Interpretation: financially constrained firms extrapolate own cost pressures, weakening the inflation expectations channel.
- Sentiment (non-political sentiment, standardized at sectoral level):
  - Empirical findings:
    - Firms one standard deviation more optimistic than sector average decrease inflation expectations by about 0.13 units in response to a contractionary monetary policy shock.
    - Amplification of the shock equal to 33%.
  - Interpretation: optimistic firms believe monetary policy is more effective; negative confidence shocks can impair effectiveness.
- Firm size (total assets, standardized at sectoral level):
  - Empirical findings:
    - Larger firms decrease their inflation expectations by about 0.07 units more after a contractionary monetary policy shock.
    - Amplification equal to 17%.
  - Interpretation: larger firms better monitor macro shocks and transmission.
- Attention to the central bank (Attention standardized at sectoral level):
  - Correlations:
    - Attention index vs actual inflation = 0.577
    - Attention index vs Fed Funds Rate = 0.792
  - Empirical findings:
    - Firms paying one standard deviation more attention than sector average decrease inflation expectations by about 0.04 units more in response to a contractionary monetary policy shock.
    - Average firm response peaks at 0.4 units of the index four quarters after the shock.
    - Amplification effect for one standard deviation more attention ≈ 10%.
  - Robustness: results robust to absence of time fixed effects, controls for information effect, and interactions with firm characteristics (Appendix B.2, B.4, B.5).

### Disagreement regimes and state-dependent effects
- Regime-switching specification based on sector disagreement index z_{s,t−1} and smooth transition function F(z_{s,t−1}) = exp(θ z_{s,t−1}) / (1 + exp(θ z_{s,t−1})), with θ calibrated to 3 (robustness to θ reported).
- Key findings (Appendix B.6 and Figure 18):
  - Low disagreement regime: firms lower inflation expectations by around 0.5 index units after a one standard deviation contractionary monetary policy shock.
  - High disagreement regime: initial decrease in expectations is followed by a rebound; overall influence of monetary policy on firms’ inflation expectations is less pronounced.
  - Result corroborates aggregate findings that monetary policy transmission is state-dependent on disagreement.
- Attention × disagreement interaction (Figure 20):
  - Firms one standard deviation more attentive than sector average lower expectations more than peers during low disagreement.
  - When disagreement is high, this differential effect becomes insignificant.
  - Correlation between firms’ inflation expectation disagreement and firms’ attention to the central bank: ρ = 0.63.
- Forecasting precision test:
  - Median mean square error of less attentive firms is 3.9% higher than more attentive firms during low disagreement.
  - During high disagreement, the forecasting advantage of attentive firms diminishes to approximately zero.
  - Interpretation: attention improves information and forecasting precision in normal times but is ineffective during high disagreement, consistent with deteriorated information quality.

### Policy implications and interpretations
- The inflation expectations channel of monetary policy:
  - May be weaker in economies with higher private debt because financially constrained firms extrapolate own cost increases.
  - Is amplified by firms’ attention to the central bank (about 10% amplification for one standard deviation higher attention).
  - Is larger for firms with more optimistic sentiment and for larger firms.
- Central bank communication:
  - Improving agents’ attention and understanding of transmission mechanisms can enhance monetary policy effectiveness.
  - In periods of high disagreement, central banks should focus communication on reducing disagreement to mitigate output losses from disinflation and to preserve the expectations channel.
- Information quality considerations:
  - During high-disagreement periods, additional attention may yield null differential benefits either because attention distributions shift right (less cross-sectional variation) or because information quality deteriorates (volatile signals or unclear communication).

*Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023202-print-pdf.pdf*

### References .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .

### References

### Introduction and motivation
- Inflation reached decades-high levels in 2022 and persistent core inflation; households’ and firms’ inflation expectations can be crucial since they are primary price and wage setters.
- Households anticipating higher inflation may demand higher wages and firms may increase prices, potentially intensifying inflationary pressures.
- Households’ and firms’ inflation expectations can differ from market participants’ and professional forecasters’ expectations due to more limited information sets and different biases.
- Figure 1 shows US professional forecasters, market-based, SoFIE survey-based firms and households’ 12-month ahead inflation expectations and highlights differences in timing and magnitude across agents.

### Novel data and methodology: ECFIE index
- Construct an Earnings Calls-based Firm’s Inflation Expectations (ECFIE) index using text-mining and machine learning applied to earnings call transcripts; index proxies firms’ inflation expectations based on the intensity of firms’ discussion about inflation.
- Earnings calls are quarterly calls between management and stakeholders that provide real-time views on the company’s outlook and are publicly available, enabling replication and alignment with balance sheet data via Compustat.
- The methodology yields an aggregate index and allows sector- and firm-level analysis.

### Coverage and frequency advantages over surveys
- Traditional firm surveys are scarce, time-consuming, and have limited firm coverage; desirable research-quality surveys should be high frequency (monthly or quarterly) and have a sample size with an average of more than 350 responses.
- Our sample consists of more than one thousand firms since 2002 and more than two thousand firms since 2004.
- The ECFIE index encompasses 96.3% of the US market capitalization.
- The ECFIE index addresses time coverage limitations (it started in 2002) and avoids priming issues tied to survey question wording by using a consistent textual-extraction methodology.

### Validation and predictive power
- The ECFIE index is highly correlated with survey data on firms’ inflation expectations and has predictive power for future inflation.
- Most analysis is conducted for the United States, with some validation exercises extended to a larger sample of advanced and emerging market countries.
- The paper performs external validation tests to ensure the index properly captures firms’ inflation expectations.

### Findings on expectations formation and deviations from FIRE
- The paper documents several violations of the full-information rational expectations (FIRE) assumption in firms’ expectations formation.
- Firm-level growth sentiment and firm characteristics, such as leverage, influence firms’ economy-wide inflation expectations and their perceived effect of monetary policy on inflation.
- High sales growth firms are more optimistic about disinflation after a monetary policy shock, suggesting the inflation expectations channel may be impaired during periods of negative confidence shocks.
- Highly leveraged firms decrease their inflation expectations relatively less following a monetary policy shock, indicating firms may extrapolate their own perceived inflation when forming aggregate inflation expectations; this implies the inflation expectations channel of monetary policy could be weaker in economies with higher private debt.

### Attention to the central bank and monetary policy transmission
- Construct an Earnings Calls-based Attention to the Central Bank index using the same methodology as the ECFIE index; earnings call intensity naturally captures attention.
- Results suggest firms’ attention to the central bank amplifies the impact of monetary policy shocks on inflation expectations by 10 percent.
- This amplification implies monetary policy effectiveness can be enhanced by strategies that improve agents’ attention and understanding of the transmission mechanism of central bank actions.
- The amplification mechanism seems to fail in periods of high disagreement, when predictive power of firms decreases and forecast errors increase, indicating potential shortcomings in central bank communication during high uncertainty.

### Paper organization (as provided)
- Section I discusses existing firms’ inflation expectations surveys and related literature.
- Section II presents the data.
- Section III describes the natural language processing methodology.
- Section IV shows properties of the new index and external validation tests.
- Sections V and VI present empirical analysis on the expectation channel of monetary policy and the role of firm attention to the central bank.
- Section VII concludes.

*Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023202-print-pdf.pdf*

### 1.2    Related literature

### 1.2    Related literature

### Contribution to text-as-data literature
- Positions the paper within growing literature applying language processing to extract data from text (Gentzkow, Kelly, and Taddy (2019)).
- Notes traditional text sources: newspapers (Ferreira and others (2019), Baker, Davis, and Levy (2022) and Caldara and Iacoviello (2022)), central bank statements (Hansen and McMahon (2016) and Handlan (2020)), tweets (Baker and others (2021)), and earnings call transcripts.
- Highlights prior uses of earnings call transcripts to measure firm-level phenomena: offshore sale of output (Hoberg and Moon (2017)), political and non-political risk (Hassan and others (2019)), attention to monetary policy (Song and Stern (2020)), exposure to epidemic diseases (Hassan and others (2021a)), and tax policy expectations (Gallemore and others (2021)).
- Claims novelty: "To the best of our knowledge, our study represents the first attempt to construct a cross-country and firm level inflation expectations index using advanced text mining techniques."

### Expectations formation and FIRE (Full-Information Rational Expectations)
- Reviews literature testing FIRE and expectation formation:
  - Coibion and Gorodnichenko (2015) tested and rejected FIRE using professional forecasters.
  - Weber, Gorodnichenko, and Coibion (2023) and Kamdar and others (2018) documented strong positive correlation between households’ expectations of unemployment and inflation.
  - Andrade and others (2022) found firms’ aggregate expectations respond to industry shocks without aggregate effects, violating FIRE.
  - Song and Stern (2020) documents heterogeneity in firms' attention to macro variables affecting responses to monetary policy.
- Paper's threefold contribution on expectation formation:
  - Shows firms’ expectations do not perform according to the FIRE assumption: firms’ characteristics and sentiment impact how firms adjust expectations after a monetary policy shock.
  - Documents extrapolation: firms extrapolate their own costs when formulating inflation expectations for the aggregate economy.
  - Financially constrained firms (more exposed to interest rate increases due to reliance on external finance) increase their inflation expectations relatively more than unconstrained firms after a monetary policy shock.
  - Firms with more positive "sentiment" (using Hassan and others (2019) non-political sentiment indicator) tend to be relatively more optimistic about disinflation following a contractionary monetary policy shock.
  - Constructs a firm-level attention to monetary policy index at higher frequency than prior work; finds the inflation expectations channel of monetary policy is larger in firms that pay more attention to monetary policy — suggesting awareness and information can improve monetary policy transmission and give central banks room to leverage this channel.

### Disagreement and monetary policy transmission
- Situates contribution in literature on disagreement among agents:
  - Stylized facts of disagreement across firms, professional forecasters, and households with substantial time variation (Mankiw, Reis, and Wolfers (2003); Dovern, Fritsche, and Slacalek (2012); Andrade and others (2016); Reis and others (2020)).
  - Evidence that disagreement affects monetary policy transmission (Falck, Hoffmann, and Hür tgen (2021); Esady (2022)).
- Paper’s contributions on disagreement:
  - Develops a new measure of firms’ inflation expectation disagreement with the novel index.
  - Using micro data, corroborates Falck, Hoffmann, and Hür tgen (2021): when disagreement is high, monetary policy leads to a rise in inflation expectations.
  - Finds that when disagreement is high, monetary policy loses additional effectiveness via the expectations channel for firms paying more attention to the central bank relative to other firms.
  - Policy implication: in periods of high disagreement, central bank communication should focus on reducing disagreement to mitigate output losses from disinflation (see Esady (2022)).

### Data overview
- Training sample and cross-country coverage:
  - Collected from the S&P Capital IQ 13,407 US companies earnings call transcripts released in 2022 to build the training sample.
  - Index computed using more than 200,000 transcripts of publicly listed companies over 39 countries, of which 26 are advanced economies and 13 emerging markets.
  - Most firms hold at least one earnings call per quarter, yielding approximately four observations per year at the firm level.
  - Actual computation performed using NL Analytics.
- Additional firm-level data:
  - Merges firm inflation expectation index with quarterly Compustat balance sheet data for publicly listed firms.
  - Controls include total assets, sales growth, current assets as a share of total assets, employment, and leverage.
- Firms non-political sentiment:
  - Uses Hassan and others (2019) firm-level non-political sentiment dataset to assess sentiment’s impact on monetary policy transmission.
- Countries with more than 500 transcripts available (listed in source): Argentina, Australia, Austria, Belgium, Bermuda, Brazil, Canada, Chile, China, Colombia, Denmark, Finland, France, Germany, Greece, Hong Kong, India, Ireland, Israel, Italy, Japan, Mexico, Netherlands, New Zealand, Norway, Poland, Portugal, Russia, Singapore, South Africa, South Korea, Spain, Sweden, Switzerland, Thailand, Turkey, United Kingdom, United States.

### Monetary policy shocks
- Adopts two high-frequency identification measures:
  - Gürkaynak, Sack, and Swanson (2005) and Nakamura and Steinsson (2018).
- Measurement specifics:
  - Shocks measured as changes in the fed funds futures rate during a specific one-hour period surrounding FOMC announcements.
  - Convert high-frequency shocks to quarterly frequency through time aggregation following Ottonello and Winberry (2020): construct a moving average of raw shocks, each shock weighted by number of days in the quarter after it occurs to weigh shocks by how long firms have had to react.
  - Main analysis uses Gürkaynak, Sack, and Swanson (2005) shock; Nakamura and Steinsson (2018) shock used in Appendix B.3 as robustness test.
- Note: "The monetary policy shocks series were extended until 2022.Q3 by Acosta (2022) and can be found in the author’s website."

### Methods: Building the indexes with NLP
- Methodological framing:
  - Situates method within literature using "bag of words" and more recent deep learning sentence-level classification (BERT and descendants) for sentiment analysis.
  - Paper uses a novel approach combining Generative Pre-trained Transformers (GPT) and human judgment to select dictionary words for index construction.
- ECFIE (Earnings Calls-based Firm’s Inflation Expectations Index):
  - Underpinning idea: intensity of discussion about future inflation in earnings calls proxies firms’ inflation expectations (the more managers talk about future inflation, the more they are concerned, indicating rising expected inflation).
  - Two main steps for dictionary construction:
    1. Selected the top 100 transcripts with the highest frequency of two set of words "inflation" and "expectation" among all the 2022 earnings call transcripts in the USA. From these transcripts, randomly selected 4000 sentences to be the training sample.
    2. Two people used human judgment to classify sentences as referring to inflation expectations or not. Fed the same sentences to the GPT model and performed machine classification. If human and machine agreed a sentence referred to inflation expectations, GPT was asked which words were key for this classification. The resulting list of words make up the dictionary fed to NL Analytics to build the ECFIE index.
  - Index calculation (rescaled for presentation):
    - ECFIE Index_it = 1000 × (P Sentences with Inflation Expectations_it) / (P Sentences_it)
    - Stated verbally: "The index is calculated as the number of sentences with two set of words 'inflation' and 'expectation' in our dictionary divided by the total number of sentences in the transcript (the index is re-scaled for presentation purposes)."
  - Applied same method to all transcripts in the US and other countries to build a cross-country time series index.
  - Reports methodological superiority over a "naive" approach:
    - Table 1 correlations (method vs Naive Search):
      - Livingston Survey: 0.8415 vs 0.7545
      - CPI: 0.8096 vs 0.6707
    - Notes increases of approximately 0.09 and 0.14 in correlation with Livingston Survey and CPI respectively relative to naive approach.
- ECFACB (Earnings Calls-based Firm’s Attention to the Central Bank Index):
  - Constructs firm-level attention to the central bank via a similar methodology adapted from Song and Stern (2020).
  - Differences from Song and Stern (2020):
    - Applies index to firms’ earnings calls in US at a quarterly frequency instead of US firms’ 10-K filings at a yearly frequency.
    - Uses an extended dictionary to better capture attention to the Federal Reserve and monetary policy (Appendix A.1 reports the list of keywords).
    - Uses sentence-level classification instead of word-level.
  - Attention index formula:
    - Attention to the FED Index_it = 1000 × (P Sentences with Central Bank words_it) / (P Sentences_it)
  - The attention index reflects frequency of sentences discussing monetary policy in earnings call transcripts, providing firm-level information on attention to the central bank.

### Key methodological and empirical claims
- Human+GPT-guided dictionary selection improves correlation with external measures of expectations (Livingston Survey, CPI, SoFIE) relative to naive word selection.
- High-frequency, sentence-level measures permit quarterly firm-level tracking of inflation expectations and attention to monetary policy.
- Empirical findings summarized earlier: heterogeneity in response to monetary policy shocks by firm financial constraint, sentiment, and attention; role of disagreement in weakening expectations channel; communication recommendation for central banks in high-disagreement periods.

*Source: wpiea2023202-print-pdf - 1.2    Related literature*

### section 4, we discuss the validation of our attention index and in Section 6, we use it to study how

### wpiea2023202-print-pdf - section 4, we discuss the validation of our attention index and in Section 6, we use it to study how

### External validation tests of our indexes: firm’s inflation expectations and attention to the central bank
- Validation primarily focused on the US; some exercises extended to other countries in the appendix.
- First exercise: correlation of ECFIE index with aggregate inflation and survey-based measures (Livingston Survey).
  - Correlation of ECFIE index with Livingston Survey: 0.84
  - Correlation of ECFIE index with CPI-U inflation: 0.8
  - Appendix A.2: shows strong positive correlations between the index and firms’ surveys for both advanced economies and emerging market economies.
- Time-dimension via sample cross-correlation function (CCF) with Livingston Survey at different horizons:
  - Highest correlation with short-term inflation (earnings calls target near-term).
  - Table 3 correlations with Livingston Survey horizons:
    - Next 1M (end of the period): 0.8296
    - Next 6M (end of the period): 0.8339
    - Next 12M (end of the period): 0.8250
    - Following 1Y (average): 0.8415
    - Next 10Y (average): 0.1383

### Comparison with other survey indexes
- Comparison with SoFIE:
  - Correlation between ECFIE and SoFIE: ρ = 0.9764
  - Disagreement correlation between the two expectation measures since survey inception: ρ = 0.7996
- Figure 3 and Table 2 summary (correlations table):
  - ECFIE index vs Livingston Survey: 0.84
  - ECFIE index vs CPI: 0.8
  - Livingston Survey vs CPI: 0.83

### Sectoral validation using equity market responses
- Method: identify sectors exposed to inflation surprises via sectoral returns after inflation releases, then check which sectors respond to changes in ECFIE.
- Finding: almost perfect relationship between sectors exposed to inflation surprises and sectors affected by the ECFIE index (Figure 5).
  - Interpretation: index captures inflation outlook at the sectoral level.

### Firm-level validation and response to industry shocks
- Motivation: test whether the index reflects firms’ adjustments to industry-level shocks that have no aggregate effects (à la Andrade and others (2022)).
- Local projections specification jointly estimates dynamic response of ECFIE to industry and aggregate conditions (aggregate inflation and industry inflation used).
- Result: in response to industry-level shocks with no aggregate effects, firms’ aggregate inflation expectations respond persistently (red line in Figure 6).
  - This mirrors documented “irrational” behavior where firms use local prices to infer aggregate conditions.

### Comparison across agent-types and dynamics during crises
- Recompiled cross-agent comparison replacing SoFIE with ECFIE:
  - All agent-type expectation measures transformed into z-scores for comparability.
  - Common patterns: decrease during the financial crisis and increase in 2021.
  - Differences in intensity:
    - Firm index decreased considerably less than professional forecasters’ market-based expectations in 2008.
    - Firm index increased at the peak of 2022.
  - Implication: different agents have differing intensity responses; the firm index can provide relevant information for policymakers.

### Predictive power for future inflation
- Local projection specification estimated:
  - π_{t+h} = α_h + β_H π^e_t + ξ_H X_t + ε_{t+h}
  - Dependent variable: year-on-year CPI inflation.
  - π^e_t: firm’s inflation expectations index (ECFIE).
  - Controls X_t: four lags of CPI, GDP growth and unemployment rate.
- Key parameter β_H plotted in Figure 8:
  - An one-unit increase in ECFIE is associated with an increase in two percent inflation in the US in the first four quarters.
  - Conclusion: the firm’s indicator has predictive power for future inflation.

### Validation of attention to the central bank (ECFACB)
- ECFACB measures intensity of firms talking about the central bank using a dictionary: Fed, Fed Funds, monetary policy, central bank, FOMC, monetary policy, quantitative easing, quantitative tightening, monetary easing, monetary tightening, Federal Reserve.
- External validation via Google Trends using the same keywords:
  - Correlation between ECFACB and Google Trends measure: 0.557
  - Interpretation: correlation suggests index correctly captures attention to the central bank and monetary policy, though measures are not perfectly collinear (Figure 9) — reflecting differences in attention between firms and the general public.

### Stylized facts: expectations by firm size and leverage
- Firms categorized into percentiles by total assets: 0–33rd, 33rd–67th, 67th–100th.
- Findings on firm size:
  - Larger firms (by total assets) discuss inflation expectations more than smaller firms.
  - Despite level differences, high correlation across size-percentile series indicates agreement on direction of changes in inflation.
- Findings on leverage:
  - Firms with higher leverage tend to discuss inflation expectations more.
  - Suggests inflation outcomes are relatively more important for higher-debt firms.
  - Despite level differences by leverage, series are highly correlated.

*Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023202-print-pdf.pdf*

### 5.2    Deviation from FIRE assumption

### 5.2    Deviation from FIRE assumption

### Validation and relevance of the aggregated inflation expectations index
- The aggregated inflation expectations index has predictive power for future inflation and correlates well with other inflation expectations surveys and available disagreement measures.
- The index exhibits characteristics analyzed in studies of firm-level inflation expectations, supporting its use to study the inflation expectations channel of monetary policy.
- Improved understanding of this channel can help central banks better target communications to help the public understand the monetary policy stance and the economic outlook.

### Sectoral heterogeneity in inflation expectations
- Inflation expectations depend on sector-specific shocks.
- Sectors represented by dashed lines (Consumer Non-Cyclical, Consumer Cyclical, Basic Materials, Industrials and Energy):
  - Had higher inflation expectations during the recent inflationary episode.
  - Reached peak expectations in 2022Q2.
- Sectors characterized by solid lines (Utilities, Financials, Real Estate, Healthcare and Technology):
  - Had lower inflation expectations.
  - Experienced peaks in either 2022Q3 or 2022Q4.
- When normalizing by sector average and standard deviation, the magnitude of the recent increase in inflation expectations was similar across sectors but the timing differed.
- It took longer for the inflation expectations index to pick up and peak in sectors that usually don’t pay attention to inflation (e.g., the health sector), suggesting heterogeneity in how firms incorporate recent information.
- These patterns align with Andrade and others (2022) and the validation exercise in Section 4 indicating firms adjust aggregate inflation expectations in reaction to industry-level shocks.

### Firms’ characteristics and the monetary policy transmission
- If firms followed the full-information rational expectations (FIRE) model, they would have similar views about how monetary policy impacts inflation regardless of their own outlook. The data shows deviations from this assumption.
- The matched micro data (firm expectations with balance-sheet data) allows testing how firm-level attributes affect responsiveness to monetary policy shocks and yields implications for central bank communication.

### Financial constraint and monetary policy effectiveness
- Hypothesis: Firms extrapolate from their own costs when forming aggregate inflation expectations; financial constraint proxied by long-term debt maturing within one year is used.
- Estimated specification (Equation 2) includes interaction FinancialConstraint_{j,t−1} × MP_t; FinancialConstraint is standardized at the sectoral level for each period (value equal one = firm is one standard deviation more financially constrained than the average firm in the sector at period t).
- Empirical findings:
  - Firms facing higher financial constraints tend to increase their inflation expectations by approximately 0.15 units (measured by the index) more in response to monetary policy shocks.
  - The average impact of a monetary policy shock results in a decrease of 0.4 units in the expectation index.
  - For a firm one standard deviation more financially constrained than the sector average just before the shock, the total effect of the shock is weakened by 37%.
- Interpretation: The inflation expectations channel of monetary policy might be notably weaker for firms with a larger portion of debt maturing within one year because increased debt-service exposure leads them to rely more on perceived own-cost increases.

### Sentiment and monetary policy effectiveness
- Hypothesis: Firms’ sentiment influences how they form aggregate inflation expectations, using a non-political sentiment (NPSentiment) measure standardized at sectoral level (value equal one = firm is one standard deviation more optimistic than sector average).
- Estimated specification (Equation 3) includes interaction NPSentiment_{j,t−1} × MP_t.
- Empirical findings:
  - Firms with a more positive outlook than their sector decrease their inflation expectations by about 0.13 units in response to a contractionary monetary policy shock.
  - This represents an amplification of the monetary policy shock equal to 33%.
- Interpretation: More optimistic firms believe monetary policy is more effective in lowering inflation; negative confidence shocks could reduce monetary policy effectiveness. This aligns with Tenreyro and Thwaites (2016) that monetary policy shocks are less effective in recessions.

### Firm size and monetary policy effectiveness
- Hypothesis: Larger firms may react differently to monetary policy shocks; Size measured by total assets standardized at the sectoral level (value equal one = firm is one standard deviation larger than sector average).
- Estimated specification (Equation 4) includes interaction Size_{j,t−1} × MP_t.
- Empirical findings:
  - Larger firms decrease their inflation expectations by about 0.07 units more after a contractionary monetary policy shock.
  - This represents an amplification of the monetary policy shock equal to 17%.
- Interpretation: Larger firms may have more resources to monitor macroeconomic shocks and better understand monetary policy transmission.

### Attention to the central bank: features and implications
- Attention index shows large heterogeneity across sectors:
  - Financial sector pays more attention to the central bank.
  - Real estate sector increased attention after the Global Financial Crisis.
  - Attention increased across most sectors during the recent tightening cycle.
- Correlations:
  - Correlation between attention index and actual inflation = 0.577.
  - Correlation between attention index and Fed Funds Rate = 0.792.
- Local projection specification for attention (Equation 5) uses Attention_{j,t−1} × MP_t; Attention is standardized at sectoral level (value equal one = firm pays one standard deviation more attention than sector average).
- Empirical findings:
  - Firms paying one standard deviation more attention than the sector average decrease their inflation expectations by about 0.04 units more in response to a contractionary monetary policy shock.
  - Average firm response to a contractionary monetary policy shock is persistent and peaks at 0.4 units of the index four quarters after the shock.
  - The amplification effect of paying one standard deviation more attention than the sector average is around 10%.
- Robustness:
  - Appendix B.2 estimates the dynamic version without time fixed effects.
  - Appendix B.4 controls for the information effect of monetary policy announcements.
  - Appendix B.5 shows impulse responses controlling for interactions between firm-level characteristics and monetary policy; results are robust.

### Firms’ disagreement and monetary policy effectiveness
- Strategy follows Falck, Hoffmann, and Hürtgen (2021): estimate firms’ inflation expectation responses π^e_{j,t+h} to a monetary policy shock MP_t conditioned on probability of being in a high- or low-disagreement regime using local projections (h = 0,...,7).
- Regime-switching specification (Equation 6) uses a smooth transition function F(z_{s,t−1}) = exp(θ z_{s,t−1}) / (1 + exp(θ z_{s,t−1})), with z normalized to unit variance of disagreement and θ calibrated to 3.
- The probability F(z_{s,t−1}) reflects the sector’s probability of being in a high-disagreement regime at t−1; z_{s,t−1} is lagged by one period to mitigate endogeneity.
- The specification allows conditioning monetary policy effects on the likelihood of high versus low disagreement regimes and yields regime-specific effects β^H_h and β^L_h. Appendix B.6 reports aggregate-level disagreement regime and robustness.

*Source: wpiea2023202-print-pdf - 5.2    Deviation from FIRE assumption*

### Appendix B.6 show that the results are robust toθvalue.

### Appendix B.6 show that the results are robust toθvalue.

### Firms’ disagreement about inflation expectations and monetary policy effectiveness
- Figure 18 illustrates Impulse Response Functions (IRFs) for the US firms’ inflation expectation index after a one standard deviation contractionary monetary shock, conditional on sector disagreement regime.
  - Confidence intervals are set at 68% and 90%.
  - Standard errors are two-way clustered by firms and time.
  - The horizontal axis shows the impulse-response horizon measured in quarter.
- Key empirical findings from Figure 18:
  - In a low disagreement regime, firms lower their inflation expectations by around 0.5 index units after a contractionary monetary policy shock.
  - In a high disagreement regime, the influence of monetary policy on firms’ inflation expectations is less pronounced: expectations initially decrease but then rebound.
  - These firm-level results corroborate aggregate findings in Falck, Hoffmann, and H ̈urtgen (2021) and support state dependence of monetary policy on the disagreement regime.
- Interpreted mechanism (literature motivation):
  - In the high-disagreement regime, firms may use central bank interest rate decisions as a signal of supply and demand conditions; an unexpected increase in interest rates can be perceived as a signal that demand is increasing, inducing firms to raise prices.

### Firms’ attention, disagreement and monetary policy effectiveness
- Subsection 6.2 result (referenced):
  - Firms paying more attention than their peers to the central bank right before the monetary policy shock decrease their inflation expectation by 10% more than the average firm in the sector, enhancing monetary policy effectiveness.
- Theoretical context and open question:
  - Rational inattention models (Máckowiak and Wiederholt (2009); Zhang (2017)) imply firms focus on idiosyncratic shocks and may respond slowly to monetary policy unless aggregate volatility rises, which shifts attention to aggregate conditions and can reduce real effects of nominal shocks.
  - Increased perceived uncertainty can motivate greater information capacity and allocation to aggregate conditions, improving detection and response to aggregate shocks.
- Correlation evidence:
  - Figure 19 shows a high correlation between firms’ inflation expectation disagreement and firms’ attention to the central bank: ρ= 0.63.
  - This suggests during periods of disagreement about aggregate inflation, firms increase information capacity and allocate more resources to monitor the central bank.

### State-dependent amplification of attention to the central bank
- Estimation approach:
  - A local projection is estimated (equation displayed in the source) with regime-dependent terms: coefficients of interest are βLh and βHh interacting Attentionj,t−1 with MPt, and F(zs,t−1) gives the probability of the sector being in a high-disagreement regime.
  - Firm fixed effects αj and time fixed effects αt are included; controls are the same used in Equation 5.
- Results (Figure 20):
  - Firms one standard deviation more attentive to the central bank than the average firm in the sector one period before the shock lower their inflation expectations more than other firms during times of low disagreement.
  - When disagreement is high, this differential effect becomes insignificant across almost all time periods.
- Interpretation:
  - Consistent with rational inattention with endogenous information choice: in high-disagreement periods, sector attention distributions shift right and are more centered, so additional attention yields null differential benefits.
  - Alternative hypothesis: information quality may deteriorate during high disagreement because economic signals become too volatile or central bank guidance/communication becomes less transparent, reducing the benefit of attention.

### Testing information quality: forecasting precision across regimes
- Design:
  - Compare forecasting power of firms’ inflation expectations index for firms that pay more attention than the median of their sectors across two environments:
    1. Low disagreement and low inflation.
    2. High disagreement and low inflation.
  - Regression of each firm’s future inflation one year ahead on firms’ inflation expectations up until two sample cutoffs: the fourth quarter of 2014 and the second quarter of 2017.
  - From 2015Q1 until 2017Q2: low inflation and low disagreement. From 2017Q3 until 2019Q4: low inflation and high disagreement.
  - Compute mean squared errors ten periods ahead for each sector, for high- and low-attention firm groups; for each sector and disagreement regime calculate the ratio of average mean square error of low-attention firms to high-attention firms; report the simple median of these ratios across sectors.
- Empirical result:
  - The median mean square error of less attentive firms is 3.9% higher than that of more attentive firms during periods of low disagreement.
  - During periods of high disagreement, this difference diminishes to approximately zero.
- Implication:
  - Attention to the central bank improves information and forecasting precision in normal times but is ineffective during periods of high disagreement, consistent with deteriorated information quality in those periods.

### Conclusion (from the broader chapter content)
- The paper builds quantitative measures of firms’ inflation expectations using text-mining of earnings call transcripts and validates the index with multiple tests.
- In the US application:
  - The index captures firms’ inflation expectations and has predictive power for future inflation.
  - Matching firm-level expectations with balance sheet data reveals violations of the full-information rational expectations (FIRE) assumption.
  - The inflation expectations channel of monetary policy may be weaker in economies with higher private debt and in periods of negative confidence shocks.
  - Monetary policy is more effective in shaping expectations when firms are more attentive to the central bank, suggesting effectiveness can be enhanced by strategies that improve agents’ attention and understanding of transmission mechanisms.
  - The inflation expectations channel appears impaired during periods of high inflation expectations disagreement, with potential implications for central bank communication strategy in periods of uncertainty.

*Source: wpiea2023202-print-pdf - Appendix B.6 show that the results are robust toθvalue.*

### Appendix A    Additional Validation

### Appendix A    Additional Validation

### Appendix A.1 — Keywords
- Appendix Table A.1 lists dictionary keywords used in the constructed text-based inflation expectation index.
  - Inflation dictionary keywords: cogs inflation, commodities inflation, core inflation, cost inflation, gross inflation, inflation, inflationary, inflationary environment, inflationary pressures, input cost inflation, market inflation, price increases, wage inflation
  - Expectations dictionary keywords: additional, ahead, concern, continue, continued, evolution, expect, expectations, expected, first quarter, fiscal year, forecast, forecasting, forward looking, forward, full year, further, future, long term, medium term, near term, next year, next years, outlook, possibility, possible, potential, pressure, projected, projections, q4, second quarter, short term, trends, year progresse.
- Notes: Keywords were selected based on the methodology described in Section 3.

### Appendix A.2 — Validation Cross-Country
- Appendix Figure A.1: Correlations between the novel firms’ inflation expectation index (blue, RHS) and existing firms surveys (red, LHS) for:
  - United Kingdom, Italy, Mexico, Sweden, Japan, Norway, New Zealand, and Turkey.
- Source networks used for comparison: S&P Capital IQ, NL Analytic, Bank of England, Italy Central Bank, Mexico Central Bank, Sweden Central Bank, Japan Central Bank, Norges Bank, Reserve Bank of New Zealand, Turkey Central Bank, and authors’ calculations.

### Appendix Table A.2 — Keyword selection method matters (SoFIE)
- Correlations:
  - ECFIE Index: 0.9764
  - Naive Search: 0.9570
- Source: S&P Capital IQ, NL Analytic, Cleveland Bank, and authors’ calculation.

### Appendix Table A.3 — Keyword selection method matters (SoFIE Disagreement)
- Correlations Disagreement:
  - ECFIE Index: 0.7996
  - Naive Search: 0.7574
- Source: S&P Capital IQ, NL Analytic, Cleveland Bank, and authors’ calculation.

---

### Appendix B    Additional Empirical Results

### Appendix B.1 — US statistics
- Appendix Figure B.2: Number of US companies used to construct the inflation expectation index over time (quarterly transcripts available for US-based companies).
- Appendix Table B.4: Summary statistics of firms’ inflation expectations and firms’ characteristics (exact values):
  - All Firms (Obs 178479)
    - Inflation Expectation Index: Mean 0.882, SD 2.98, Median 0.00
    - Total Assets (US$million): Mean 1784791865.52, SD 6822.54, Median 3444.98
    - Leverage: Obs 173691, Mean 0.27, SD 0.34, Median 0.23
  - High Inflation Expectations (Obs 89220)
    - Inflation Expectation Index: Mean 1.664, SD 4.01, Median 0.00
    - Total Assets (US$million): Mean 892202176.54, SD 6183.01, Median 544.33
    - Leverage: Obs 87399, Mean 0.29, SD 0.24, Median 0.27
  - Low Inflation Expectations (Obs 89259)
    - Inflation Expectation Index: Mean 0.100, SD 0.65, Median 0.00
    - Total Assets (US$million): Mean 892591554.63, SD 7393.78, Median 207.96
    - Leverage: Obs 86292, Mean 0.24, SD 0.41, Median 0.16
- Notes: Leverage defined as ratio of total debt to total assets. Firms split by whether average inflation expectation is above or below the sample median.

- Appendix Figures B.3 and B.4:
  - B.3: Evolution of inflation expectation by economic sector for the entire sample from 2002.Q1 until 2023.Q1.
  - B.4: Evolution of inflation expectation by sector standardized to mean zero and standard deviation equal to one from 2020.Q1 until 2023.Q.1.

### Appendix B.2 — The Average Effect of Monetary Policy
- Estimation equation (B.1) used to quantify average effect of monetary policy shocks on firms’ inflation expectations:
  - π^e_{j,t+h} = α_j + γ^H_{t+h} MP_t + δ^H_{t+h} Attention_{j,t−1} × MP_t + β_{t+h} X_{j,t−1} + η_{t+h} Y_{j,t−1} + ε_{j,t}
  - Where Y_{j,t} is a vector with four lags of GDP growth, the inflation rate, and the Fed Fund Rates; X_{j,t} is a vector of firm level controls (same used in Equation 5).
- Key finding:
  - The average response to monetary policy, γ^H, peaked after four quarters reducing the inflation expectation index by approximately 0.4 units after an one standard deviation contractionary monetary policy shock.
- Appendix Figure B.5: Impulse-response horizon measured in quarter; confidence intervals set at 68% and 90%; standard errors two-way clustered by firms and time.

### Appendix B.3 — Additional Results with an Alternative Monetary Policy Shock
- Re-estimated Equations 5, 2, 3, 4 using the monetary policy shock constructed by Nakamura and Steinsson (2018).
- Key finding:
  - Figure B.6 shows results are robust to using the Nakamura and Steinsson (2018) monetary policy shock instead of Gürkaynak, Sack, and Swanson (2005).
- Appendix Figure B.6 panels: (a) Attention, (b) Financial Constraint, (c) Non-Political Sentiment, (d) Size. Confidence intervals at 68% and 90%; standard errors two-way clustered by firms and time; horizon in quarters.

### Appendix B.4 — Robustness check for the information channel of monetary policy
- Concern: FOMC announcements may contain information about future economic activity.
- Approach: Re-estimate Equations 5, 2, 3, 4 including controls for interaction between the variable of interest (Attention_{j,t−1}, FinancialConstraint_{j,t−1}, NPSentiment_{j,t−1}, Size_{j,t−1}) and the “future path of policy” factor from Gürkaynak, Sack, and Swanson (2005).
- Key finding:
  - Figure B.7 shows results are robust to controlling for the information channel of monetary policy.
- Appendix Figure B.7 panels: (a) Attention, (b) Financial Constraint, (c) Non-Political Sentiment, (d) Size. Confidence intervals at 68% and 90%; standard errors two-way clustered by firms and time; horizon in quarters.

### Appendix B.5 — Robustness check for firm level exposure to monetary policy shock
- Concern: Disentangle firm level exposure to monetary policy from attention.
- Specification (B.2):
  - π^e_{j,t+h} = α_j + α_t + δ^H_{t+h} Attention_{j,t−1} × MP_t + β_{t+h} X_{j,t−1} + θ_{t+h} X_{j,t−1} × MP_t + ε_{j,t}
- Key result:
  - Appendix Figure B.8 shows dynamics of interaction coefficient between attention and monetary shocks over time when interacting MP with all firm level controls. Confidence intervals at 68% and 90%; standard errors two-way clustered by firms and time; horizon in quarters.

### Appendix B.6 — Robustness check for disagreement regime

- Appendix B.6.1 — Sector level
  - Concern: Sensitivity to parameter θ that determines switching between low and high disagreement regimes.
  - Approach: Estimate β^L_h and β^H_h in Equation 6 for θ = 1 and θ = 5.
  - Key result:
    - Figure B.9 shows results are robust to the value of θ.
  - Figure B.9 panels:
    - (a) Low Disagreement θ = 5; (b) High Disagreement θ = 5; (c) Low Disagreement θ = 1; (d) High Disagreement θ = 1.
    - IRFs after an one standard deviation contractionary monetary shock; confidence intervals at 68% and 90%; standard errors two-way clustered by firms and time; horizon in quarters.

- Appendix B.6.2 — Aggregate level
  - Regime-dependent specification (B.3):
    - π^e_{j,t+h} = α_j + [α^H_h + β^H_h MP_t + γ^H_h x_{j,t−1}] F(z_{t−1}) + [α^L_h + β^L_h MP_t + γ^L_h x_{j,t−1}] (1−F(z_{t−1})) + ε_{j,t+h}
    - F(z_{t−1}) = exp(θ z_{t−1}) / (1 + exp(θ z_{t−1})), where z is an index normalized to have unit variance of the disagreement.
    - S = H,L denotes high and low disagreement regimes. α_j denotes firm fixed effect.
  - Key result:
    - Figure B.10 plots β^S_h for different θ values and shows results are robust to θ.
  - Figure B.10 panels:
    - (a) Low Disagreement θ = 5; (b) High Disagreement θ = 5; (c) Low Disagreement θ = 3; (d) High Disagreement θ = 3; (e) Low Disagreement θ = 1; (f) High Disagreement θ = 1.
    - IRFs after an one standard deviation contractionary monetary shock; confidence intervals at 68% and 90%; standard errors two-way clustered by firms and time; horizon in quarters.

_ Mining the Gap: Extracting Firms’ Inflation Expectations From Earnings Calls Working Paper No. WP/2023/202_

---


_Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023202-print-pdf.pdf_