Auction Catalogue Narratives, Moral-Historical Framing, and Auction Outcomes for Adrian Ghenie Lots: Data and Code

This repository contains data, code, and replication materials for a mixed-method study of how Christie’s and Sotheby’s publicly narrate morally burdened historical references in Adrian Ghenie auction lots, and how such framing corresponds to auction outcomes.

The dataset links public, author-archived lot materials to structured lot-level variables: auction house, sale date, lot number, title, dimensions, creation year, estimates, realized prices, sale currencies, and moral-language coding fields. The corpus covers sales from approximately 2013–2025 and was collected from public online catalogues in November 2025. No private client communications, internal notes, private negotiations, or other non-public auction-house documents are included.

Moral-historical framing was coded through a transparent dictionary approach. Variables include moral_flag, moral_token_count, moral_terms_hit, and moral_intensity, an original ordinal 0–3 measure of the most intense moral-historical reference in each lot text. This measure is descriptive and study-specific; it is not a psychological scale or measure of buyer response, trauma severity, moral judgment, or the artwork’s inherent meaning.

The analytic sample contains 106 lots after excluding observations with missing/nonpositive prices or estimates, mismatched currencies for estimate-relative tests, and missing controls required for logged specifications.

The updated package reflects the revised econometric analysis. ghenie_econometrics_final.py corrects an extraction error for Sotheby’s The Sunflowers in 1937: the Ghenie lot sold for GBP 3,117,000, while the previously extracted CHF 175,000 referred to a historical Van Gogh-related catalogue reference. The script audits currency consistency, constructs the analytic sample, treats moral_flag as descriptive, and uses moral_intensity and moral_token_count as main narrative predictors.

The quantitative analysis is exploratory and associational, not causal. Main outcomes are log_premium_ratio = ln(realized price / low estimate) and premium_dummy, coded 1 when realized price exceeds the high estimate. Robustness checks include currency indicators, winsorized premium ratios, moral-flag specifications, influential-observation checks, logistic regression, and interaction models.

Results support cautious interpretation: moralized language is widespread, but continuous premium-ratio models do not show a robust moral premium. moral_intensity is positively associated with exceeding the high estimate in one exploratory binary specification, but this indicates mixed patterning, not proof that moralized catalogue language directly increases prices.