AI Exposure and the ICC–CAPM Gap: S&P 500 Data, 2020–2024
Description
This dataset supports the study “What Does the Market Price About AI? Industry Exposure, Firm Disclosure, and the ICC–CAPM Gap.” It contains a balanced panel of 250 non-financial S&P 500 firms observed annually from 2020 to 2024, for a total of 1,250 firm-year observations. The dataset combines disclosure-based measures of artificial intelligence exposure with firm-level accounting controls and point-in-time valuation variables. AI exposure is measured using keyword counts from annual reports and Form 10-K filings. Firm controls include size, return on assets, leverage, sales growth, Big Four auditor status, and CEO duality. Valuation variables include the risk-free rate, market equity risk premium, modified PEG implied cost of equity, firm implied equity risk premium, CAPM-based cost of equity, and the ICC–CAPM gap. The workbook also contains a multilevel AI decomposition for 218 firms across 53 subindustries. This decomposition separates observed AI disclosure into persistent firm exposure, leave-one-out subindustry AI orientation, firm-specific persistent AI, and annual within-firm AI variation. Pre-period AI measures for 2020–2021 and average post-period valuation outcomes for 2023–2024 are included to support temporal validation analyses. The workbook contains four sheets: the full firm-year panel, the firm-level AI decomposition, a concise README, and a variable codebook. Raw analytical values are preserved, missing observations are not imputed, and derived variables are transparently documented. Valuation inputs were obtained primarily from Bloomberg Professional Services, supplemented where necessary by reliable archival sources.
Files
Steps to reproduce
Define the sample. Select 250 non-financial S&P 500 firms with complete annual disclosure, accounting, and valuation coverage for 2020–2024, producing 1,250 firm-year observations. Collect annual filings. Retrieve each firm’s annual report or Form 10-K. Use the principal searchable document and exclude exhibits, duplicate attachments, appendices, and non-searchable image content. Construct AI disclosure measures. Apply the predefined AI dictionary using case-insensitive searches, standalone matching for “AI,” normalized hyphenated variants, and specified prefix rules. Count all valid matches and calculate: AI Log = ln(1 + AI mentions) Governance-risk phrases are counted separately and used only as supplementary variables. Collect firm controls. Record total assets, net income, debt, revenue, auditor type, CEO duality, sector, and subindustry. Calculate size as the natural logarithm of total assets, ROA as net income divided by assets, leverage as debt divided by assets, and sales growth as the annual revenue change. Collect valuation inputs. For each year-end valuation date, obtain price, first- and second-forward EPS forecasts, expected dividends, the 10-year U.S. Treasury rate, the U.S. implied market ERP, and firm beta. Bloomberg is the primary source, supplemented where necessary by reliable archival sources. Calculate valuation measures. Compute the MPEG implied cost of equity only when EPS2 > EPS1. Then calculate: Firm implied ERP = MPEG ICC − risk-free rate CAPM COE = risk-free rate + beta × market ERP ICC–CAPM gap = MPEG ICC − CAPM COE Invalid MPEG observations remain blank; no artificial values are inserted. Construct the AI decomposition. Retain firms with at least one other sampled firm in the same subindustry. Calculate persistent AI exposure as the 2020–2024 firm mean, subindustry AI orientation as the leave-one-out peer mean, firm-specific persistent AI as the difference between the two, and annual within-firm AI as the yearly deviation from the firm mean. Construct temporal measures. Average AI exposure and controls over 2020–2021 and valuation outcomes over 2023–2024. Reconstructed beta is calculated as: (CAPM COE − risk-free rate) / market ERP Preserve raw data. Do not impute missing values. Winsorization, where required, should be applied only during statistical analysis and not to the deposited raw variables.
Institutions
- Ca' Foscari University of VeniceVeneto, Venice